A curated archive of Google and Microsoft search patents organized by the inventor or research lead who shipped them. Each track follows one engineer through their patent lineage so the reader can see how a research thread evolved across continuations, divisional applications, and grants. Patents are annotated with the ranking surface they touch: query understanding, ranking, retrieval, behavioral signals, knowledge graph, or AI search.
Patent tracks by inventor
- 65 Google Patents to Help You Understand How Search Engines Work! (69 patents) - A curated collection of foundational and advanced patents filed by Google.
- Steven Baker, Google Search Patents (49 patents) - 49 granted search-engine patents plus 2 published applications by Google researcher Steven D. Baker, covering query understanding (synonyms, n-gram, geographic, cross-language), query refinement, meaningful stopword detection, answer passage scoring (featured snippets), query–document similarity, list co-occurrence, and embedding-based personalized search.
- Amit Singhal, Google Search Patents (22 patents) - 22 search-engine patents by Amit Singhal, former Google SVP of Search (2000-2016), covering interleaving and result composition, query rewriting, semantic query understanding, link-quality scoring, location-aware ranking, language and country biasing, commercial-intent detection, and advertisement serving. The 8-patent meaningful-stopwords family is co-invented with Steven Baker and is documented under the Steven Baker section to avoid duplicate content.
- Prabhakar Raghavan, Google Search Patents (16 patents) - 16 search/IR patents by Prabhakar Raghavan, head of Google Search 2020-2024 (previously Yahoo VP and IBM Almaden research scientist), covering personalized PageRank, web community detection, adaptive ranking, multi-engine result merging, hierarchical taxonomy, subspace clustering, information filtering, customizable navigation, user-generated content surfacing, location-quality signals from travel patterns, and temporal metadata analysis. Spans his IBM, Yahoo, and Google research.
- Nitin Gupta, Google Search Patents (32 patents) - 32 search/SEO patents by Nitin Gupta (Google Search Quality engineer) covering query suggestions and templates, reranking and personalization, answer-generating neural networks, content-knowledge generation, the latest 2026 generative-search retrieval patent for the AI Overviews era, and a direct SEO patent on domain names and websites. The final 5 patents are co-invented with Steven Baker and are documented in his section to avoid duplicate content.
- Navneet Panda, Google Search Patents (15 patents) - 15 search-quality patents by Navneet Panda, the Google engineer whose content-quality work became the basis for the Panda algorithm update (2011). Covers site quality scoring, predicting site quality before behavioral data accumulates, ranking search results by combined relevance and site-quality signals, website duration performance, locally significant queries, query revision, and image-concept learning.
- Anand Shukla, Google Search Patents (16 patents) - 16 search/IR patents by Anand Shukla, Google Search Quality engineer, including the cutting-edge 2024 "Search with stateful chat" patent that bridges classical search to LLM-based conversational search. Also covers in-depth article surfacing, query pattern matching, query topic mapping, content channel curation, and the enhanced-search-feed family. The final 3 patents are co-invented with Steven Baker on the personalized-feed embeddings work and are documented in his section.
- Krishna Bharat, Google Search Patents (48 patents) - 48 search-engine patents by Krishna Bharat, founding inventor of Google News and creator of the Hilltop algorithm. Covers the Hilltop link-analysis algorithm, query rewriting with entity detection, semantic units in queries, query and search augmentation, document ranking by usage statistics, image selection for news search, editorial-opinion ranking, search engine coverage estimation, query-specific bookmarking, and entity-aware ad rendering. Spans 2001 to 2024.
- Srinivasan Venkatachary, Google Search Patents (25 patents) - 26 search-engineering patents by Srinivasan Venkatachary, co-inventor on the Generative Summaries for Search Results patent family that underpins AI Overviews and Search Generative Experience. Also covers enhanced search feeds based on user interests, candidate-answer-passage scoring, context-scoring adjustments for answer passages, implicit question query identification, keyword-based conversational voice search, query generation via document structural similarity, resource identification from organic and structured content, content channel curation, and the emoji classifier. Spans 2013 to 2025.
- Jeromy William Henry, Google Search Patents (25 patents) - 25 search-engine patents by Jeromy William Henry covering Knowledge Panels (the entity card next to search results), Structured Entity Information Pages (the standalone entity-detail surface), Personalized Entity Information Pages, SERP image sizing, and music-search secondary link presentation. Every patent in his corpus is SEO-relevant. Spans 2013 to 2025.
- Ramanathan V. Guha, Google Search Patents (33 patents) - 33 search-engine patents by Ramanathan V. Guha, co-founder of Schema.org and lead inventor on Google Custom Search Engine. Covers the Programmable Search Engine family, trust-based search ranking, spam detection for programmable contexts, social-signal propagation, classifying search results, customized web summaries, filtering by user annotations, contextual searching by query intersections, query identification, and the generating-tasks UI family. Spans 2003 to 2019.
- Marc Najork, Google Search Patents (25 patents) - 25 search/IR patents by Marc Najork spanning his Microsoft Research era (web crawler architecture, spam-resistant ranking, content evaluation, hyperlink databases) and his Google era (form information extraction, ML-driven semantic similarity, active learning, document activity logs for relevance training). Includes a foundational cross-listed patent with the 65 Google Patents collection on document activity logs. Spans 2005 to 2025.
- Jeffrey Dean, Google Search Patents (78 patents) - 78 search and IR-infrastructure patents by Jeffrey Dean, Google Senior Fellow and Chief Scientist, co-founder of Google Brain, and architect of MapReduce and BigTable. Includes the foundational Document Scoring families (Query Analysis, Content Update, Inception Date, Link-Based Criteria), Historical Data Patent (cross-listed with 65gp), Anchor Tag Indexing, Content Snippets via Tokenspace, near-duplicate and mirrored-host detection, topic-drift prevention, distributed ML models, and the foundational MapReduce and BigTable infrastructure that makes web-scale search possible. Spans 2000 to 2025.
- Paul Haahr, Google Search Patents (63 patents) - 63 Google search and ranking patents by Paul Haahr, long-tenured Google search ranking engineer. Covers the canonical link-quality patents (Bharat/Singhal/Haahr "Determining quality of linked documents"), document-classification ranking, node-independence anti-PBN signal, the rumored quality-measure patent (cross-listed with 65gp pat-48), the multi-stage query processing pipeline, the framework for evaluating web search scoring functions (the documentary back-stop to his SMX talk), preferred sites, reference contexts, synthetic descriptive text for images, and the multi-family query refinement and stopword-detection clusters. Co-authored with Singhal, Dean, Bharat, Cutts, Henzinger, Simon Tong, and Acharya. Spans 2008 to 2019.
- Hyung-Jin Kim, Google Search Patents (23 patents) - 23 Google search and ranking patents by Hyung-Jin Kim. Lead inventor on the Navboost / implicit-feedback ranking family (US 8,661,029 + 5 continuations through US 11,816,114 active 2023 IP), the system surfaced at the 2024 DOJ Google antitrust trial. Also covers the presentation-bias model, temporal score adjustments, document-change ranking, user cohort grouping, similar-query borrowing, variant generalized queries, propagating query classifications, and the cross-listed Quality Measure patent (with Paul Haahr, 65gp pat-48). Co-authored with Simon Tong, Noam Shazeer, Michelangelo Diligenti, Adrian Corduneanu, Henele Adams, Andrei Lopatenko. Filings 2006-2021.
- Pandu Nayak, Google Search Patents (13 patents) - 13 Google search patents by P. Pandurang Nayak, Google Chief Scientist for Search. The patent record covers the query-revision and query-understanding stack from 2003-2015: integration of multiple query-revision models, confidence-scored revision, known-highly-ranked-queries borrowing, concept-context substitution, phrase-restricted synonyms, click-driven synonym identification, acronym expansion. Co-authored with David R. Bailey, Ben Gomes, Trystan Upstill, Kedar Dhamdhere, Thomas Strohmann. His neural-era contributions (RankBrain, BERT, MUM, DOJ testimony) are intentionally unpatented and live in blog posts and trial transcripts.
- Daniel Egnor, Google Local Search Patents (20 patents) - 20 Google local-search and geographic-relevance patents by Daniel Egnor, the architect of the Google Maps / Local Pack ranking layer. Lead inventor on US 8,046,371 "Scoring local search results based on location prominence" - the patent that defines the "Prominence" factor in Google's official Local Search Ranking documentation. Also covers location-sensitivity ranking (Relevance factor, with Amit Singhal), geographical-relevance indexing, ambiguous-geographic-reference resolution, authoritative-document identification, business-listing categorization, and GIS ambiguous-search processing (with Keyhole founder John Hanke). Co-authored with Singhal, Haahr, Greenfield, John Hanke. Filings 2003-2019.
- Yossi Matias, Google SERP Features & Assistant Patents (47 patents) - 47 Google patents by Yossi Matias, Google VP Engineering. Lead inventor on the foundational People-Also-Ask patent US 9,213,748 "Generating related questions for search queries". Built Autocomplete (2004), Trends, weather/sports/dictionary OneBoxes. Covers the People-Also-Ask family, Knowledge Panel entity scaffolding (grouping related entities, central-entity identification, fact estimation, textual entity attributes, entity-quote search), the Trends/analytics layer (campaign-competitive analysis, ranking graphical visualizations), the Duplex automated-call infrastructure, the personalized chatbot and LLM-output entailment patents that underpin Google Assistant's 2024-2026 generative era, and contextual / third-party SERP-features. Co-authored with Yaniv Leviathan (Duplex), Gal Chechik, Ziv Bar-Yossef, Eyal Segalis, Avinatan Hassidim. Filings 2013-2026.
- Andrei Broder, Cross-Vendor Search Patents (25 patents) - 25 search-engine and IR patents by Andrei Broder, the cross-vendor inventor who shaped near-duplicate detection at DEC/AltaVista, co-invented the foundational CAPTCHA at Compaq, built the AltaVista web-page ranking system, authored web-page-decay signal at IBM, and shaped ad-matching/SERP-features/CAPTCHA-evolution infrastructure at Yahoo. Lead inventor on the foundational MinHash/shingling patent (US 6,349,296), the fingerprint-collision-probability technique (US 5,974,481), the document-resemblance shingling patent (US 6,230,155), and the foundational CAPTCHA patent (US 6,195,698). Spans 1999 to 2024.
- Monika Henzinger, Google Search Patents (50 patents) - 98 search and IR patents by Monika Henzinger, the first Director of Research at Google. Co-inventor on the foundational near-duplicate-page-detection patent (US 6,138,113, with Dean) and an extensive set of document-scoring families (Query Analysis, Content Update, Inception Date, Link-Based Criteria, Historical Data - all cross-listed with Dean's canonical articles). Independent contributions: the detecting-duplicate-files family with William Pugh, the AltaVista connectivity-server, the connectivity-and-content-ranking patents, document-freshness determination, semantic-distance ranking, query-semantic-information ranking, anchor-text cross-language IR, in-context searching, hypertext-browser assistant, and usage-statistics-driven document retrieval. Spans 1997 to 2017.
- Trystan Upstill, Google Site-Quality & HCU-Era Search Patents (28 patents) - 28 Google search and ranking patents by Trystan Upstill. Co-inventor on the Panda patent (US 9,135,307 = 65gp pat-39) and lead inventor on the HCU-era follow-up US App 2019/0155948 "Re-ranking resources based on categorical quality" - the patent that anchors the post-Panda site-quality lineage. Also covers authoritative-results discovery, search-result ranking and re-scoring, affiliated-domains detection, resource-attribute extraction from site address, local search and POI identification, synonyms and query rewriting, navigational-intent resources, language identification from link context, and the recent LLM-era UI components patent. Filings 2012-2026.
- Susan Dumais, Microsoft IR & Search Patents (18 patents) - 61 search and IR patents by Susan Dumais, Microsoft Technical Fellow and Gerard Salton Award winner. Co-inventor on the foundational Latent Semantic Indexing patent (US 4,839,853, Bellcore 1989) - the conceptual ancestor of every dense-embedding retrieval system. At Microsoft Research her work covers automated SERP satisfaction measurement, preference-judgment click models, activity-based-context ranking, time-aware ranking with temporal dynamics, personalized navigation and search results, per-user domain-expertise determination, web-page-change × revisitation freshness, implicit device-related query reformulation, authority ranking. Spans 1989 to 2026. Co-authors include Jaime Teevan, Eric Horvitz, Ryen White, Adam Fourney, Joshua Goodman, Eric Brill.
- Christopher Burges, Microsoft Research Learning-to-Rank Patents (16 patents) - ~30 unique US inventions (84-entry WIPO corpus) by Christopher J. C. Burges, the Microsoft Research scientist who invented the RankNet/LambdaRank/LambdaMART learning-to-rank lineage. RankNet (US 7,689,615) is the pairwise neural ranker; LambdaRank (US 7,617,164) is the arbitrary-cost extension that optimizes NDCG; LambdaMART (US 12/032,697) is the gradient-boosted ensemble that won the Yahoo LTR Challenge 2010. Also covers behavioral-variability adaptation, web-page multi-graph analysis, link-spam smoothing, margin-tree boosting, k-NN probability estimation. Spans 2001 to 2019. Co-authors include Robert Ragno (LambdaRank), Irina Matveeva, Timo Burkard.
- Eric Brill, Microsoft Research Search & NLP Patents (20 patents) - 20 Microsoft Research patents by Eric Brill, the NLP and search scientist known for query speller patents and the noisy-channel correction model. Lead inventor on US 7,254,774 "Systems and methods for improved spell checking" - the foundational noisy-channel query speller patent. Also covers string-to-string spell-correction transformations (US 7,290,209 / 7,366,983), behavioral-variability accounting in web search (US 7,743,047), popularity-data ranking, mining web search user behavior, page-biased search, cost-benefit Q&A composition, and user-intent discovery. Filings 2005-2010.
- Tomas Mikolov, Google word2vec Patents (2 patents) - 2 foundational Google patents by Tomas Mikolov covering word2vec - the continuous word-embedding learning model that defined the embeddings era. Co-invented with Kai Chen, Greg Corrado, and Jeff Dean. US 9,037,464 is the original grant; US 9,740,680 is the continuation with broader claim scope. Both patents share the same disclosure (application 13/841,640) and together constitute the entire word2vec patent family. Cross-listed conceptually with Dean's section (Dean is co-inventor and works on the broader Brain/distributed-models infrastructure).
- Noam Shazeer, Google Transformer, MoE & Search Patents (18 patents) - ~46 captured Google patents (full portfolio estimated at 120-150) by Noam Shazeer. Co-inventor on the foundational Transformer attention architecture (US 10,452,978 with Vaswani, Polosukhin, Uszkoreit, Jones, Gomez, Kaiser, Parmar), the Sparsely-Gated Mixture-of-Experts scaling approach (US 11,769,055 with Dean, Hinton, Le, Mirhoseini), the Switch Transformer (US 12,093,829), and the Navboost implicit-feedback ranking family (US 8,661,029 with Kim/Tong/Diligenti - cross-listed). Also covers LLM-in-assistant response generation, distributed tensor computations (GSPMD/Pathways), and the 2008-2015 pre-Transformer large-scale ML ranking infrastructure. Spans 2008 to 2025.
- Ashish Vaswani, Google Transformer, Attention-Based Vision & Fast Decoding Patents (6 patents) - 4 captured new canonical articles plus 2 cross-listings from the Noam Shazeer section. Lead author of the foundational Transformer patent (US 10,452,978, cross-listed). First inventor on the local-self-attention computer vision patent (US 20210390410, HaloNet) that scales attention to long sequences and images. Co-inventor on MultiModel (US 20200089755 - the unified multi-task multi-modal architecture that is the structural ancestor of MUM and Gemini), the fully-attentional computer vision system (US 20250292560, the ViT lineage), and the fast-decoding-with-discrete-latents method (EP 3732627) that makes LLM inference economically viable at search-scale latency. Spans 2017 to 2025.
- Jakob Uszkoreit, Google Transformer Architect, Universal Transformer & Vision Transformer Patents (6 patents) - 3 new canonical articles plus 3 cross-listings from the Vaswani and Shazeer sections. Co-author of "Attention Is All You Need" (US 10,452,978, cross-listed). Co-inventor on the Universal Transformer (US 10,740,433, the adaptive-depth Transformer with per-token Adaptive Computation Time), the Decomposable Attention NLI model (EMNLP 2016, the load-bearing pre-Transformer proof that attend-compare-aggregate replaces recurrence), and the Vision Transformer (ViT, ICLR 2021, the patch-tokenization architecture that anchors modern visual search and multimodal grounding). Cross-listings cover the original Transformer, MultiModel (the unified multi-task multi-modal ancestor of MUM/Gemini), and the Image Transformer. Spans 2016 to 2021.
- Simon Tong, Google Click-Driven Ranking, Population Signals & Anti-Spam Patents (9 patents) - 5 new canonical articles plus 4 cross-listings from the Kim, Shazeer, and 65 Google Patents sections. Tong is co-inventor with Marc Pearson and Sergey Brin on the related-query ranking patent (US 7,505,964), with Pearson on the population-information ranking patent (US 7,454,417), solo on country biasing (US 20040254932), with Bem/Harik/Levenberg/Shazeer on the pre-Transformer large-scale ML ranking infrastructure (US 7,222,127), and with Ghemawat/Piscitello/Cutts on the user-document-removal patent (US 8,417,697, the structural ancestor of Personal Blocklist and crowd-sourced spam signals). Cross-listings cover the Navboost implicit-feedback family, the CTR-as-ranking-factor patent, the historical-data patent, and the large-dataset-ranking continuation. Spans 2003 to 2017+.
- Michelangelo Diligenti, Google Click-Data Quality & Implicit Feedback Patents (5 patents) - 2 new canonical articles plus 3 cross-listings from the Kim and 65 Google Patents sections. Diligenti is inventor on the Detecting Click Spam patent (US 8,694,374, attribute-deviance anomaly detection feeding ranking signal) and on the Click Model That Accounts for User Intent (US 20120143789, intent-conditional click weighting). His Navboost / presentation-bias / CTR-as-ranking-factor co-inventorships (cross-listed) anchor him in the implicit-feedback ranking family. The portfolio focuses on click-data quality, the layer between raw user behavior and the ranking signal it feeds. Spans 2007 to 2019+.
- Jaime Teevan, Microsoft Personalized Search & Search-Task Patents (5 patents) - 4 new canonical articles plus 1 cross-listing from the Dumais section. Teevan is co-inventor with Dumais and Horvitz on the foundational personalized-search patent (US 7,693,818) that re-ranks global candidates against a per-user profile vector. Inventor on search-task identification (US 8,326,824, cross-session query clustering into multi-step tasks), on re-finding vs new-finding classification (US 8,756,219, stabilizing prior clicked results for return visits), and on microtask search decomposition (US 10,062,103, decomposing high-level goals into resumable micro-steps). Cross-listing covers personalized-navigation (Dumais canonical). Spans 2003 to 2018+.
- Eric Horvitz, Microsoft Decision-Theoretic Search & Attention-Aware Information Delivery Patents (3 patents) - 3 new canonical articles. Horvitz is inventor on decision-theoretic ranking under uncertainty (US 7,124,129, relevance as a probability distribution with expected-utility selection), on cost-benefit reasoning in search-result presentation (US 7,194,464, the mechanical basis for featured snippets and AI Overviews), and on attention-aware proactive information delivery (US 6,934,917, the zero-query surfaces lineage behind Discover, Now cards, and AI assistants). Spans 2002 to 2006+.
- Ryen White, Microsoft Search Behavior & Search-Trails Patents (3 patents) - 3 new canonical articles. White is inventor on search-trail / successful-path mining (US 7,904,440, ranking destinations that end successful cross-user trails), on exploratory-versus-lookup query phase classification (US 8,041,711, serving phase-tuned ranking and result formats), and on cross-session task continuation (US 9,116,996, stitching queries across sessions and devices into one continuing task). Spans 2007 to 2015+.
- Azalia Mirhoseini, Google ML Infrastructure & Chip-Design Patents (3 patents) - 2 new canonical articles plus 1 cross-listing from the Shazeer section. Mirhoseini is co-inventor on the RL-based chip floorplan placement patent (US 11,475,278, the Nature 2021 paper that designs TPU floorplans in hours instead of weeks) and on the device-placement-for-distributed-ML patent (arXiv 1706.04972, the RL controller that learns how to split large models across GPUs/TPUs to minimize training time). Both feed back into the compute-cost economics that let Google serve Transformer ranking at search-scale. Cross-listing covers the foundational MoE patent (Shazeer canonical). Spans 2017 to 2021+.
- Arnold Overwijk, Microsoft Site Topical Authority Patent (1 patents) - 1 canonical article. Overwijk is co-inventor with Li Xiong, Chuan Hu, and Junaid Ahmed on the Site Topical Authority patent (US 20210004416, the formula A(s,t) = SR(s) * C(s) * T(s,t)^2 that explicitly quantifies per-site per-topic authority by combining global site rank, click signal, and squared per-topic page share). The most precise mathematical formulation of topical authority in the modern patent literature. Conceptually paired with Procopio (per-author per-topic authority) and Agent Rank (author-level reputation, 65gp pat-47).
- Michael Procopio, Google Per-Author Per-Topic Authority Patent (1 patents) - 1 canonical article. Procopio is named inventor on the Topic Authority patent (US 8,458,196). The system computes per (author, topic) authority signatures: for each document d the author contributed to, sum (authorship% * topic_weight); aggregate across all the author's documents to produce a per-author per-topic expertise score. The mathematical quantification of "this person is an authority on this topic," structurally the basis for the E-A-T Expertise signal at the author level. Connects to Agent Rank (US 7,565,358, 65gp pat-47, the earlier Google author-reputation patent) and to Overwijk's site-topical-authority (US 20210004416) to form a layered authority model.
- Ori Allon, Google Orion Query-Refinement Patent (1 patents) - 1 canonical article. Allon is co-inventor on the Orion query-refinement patent (US 8,392,443) with Ugo Di Girolamo, Tomer Shmiel, Alexandre Petcherski, and Tzvika Hartman. The Orion algorithm originated as Allon's PhD-thesis work at UNSW and was acquired by Google in 2006. The patent generates corpus-mined query refinements: given a query, extract phrase patterns from documents that satisfy the query to produce ranked refinement candidates. The mechanism behind "Searches related to," extended SERP snippets, and the lineage of People-Also-Ask. Granted 2013.
- Corin Anderson, Google Reasonable Surfer (Foundational) Patent (2 patents) - 1 new canonical article plus implicit cross-listings to the 65 Google Patents collection. Anderson is FIRST inventor on the foundational Reasonable Surfer patent (US 7,716,225, with Jeff Dean and Alexis Battle), the 2004-filed grant that introduces non-uniform link-click probability weighted by per-link features (anchor text, font size, position, surrounding context). The continuations (US 8,117,209 in 65gp pat-45, US 9,305,099 in 2016) extend the same model. The mechanism for differentiated PageRank flow across links on the same page.
- Ashok Popat, Google OCR & Image-Text Indexing Patents (1 patents) - 1 canonical article covering Popat's translation-inspired OCR research (Popat et al., 2009 onward; US 8,675,012 patent family for selective display of OCR'd text). The methodology reframes OCR as statistical machine translation: image features as the source channel, character sequences as the target, joint decoding with channel and language models maximizes transcription quality. Powers Google Books indexing, Lens text capture, Drive PDF search, and the image-bound-text indexing pipeline behind every image and infographic on the web.
65 Google Patents to Help You Understand How Search Engines Work!
- Information Extraction from a Database (DIPRE Algorithm) (US 6,678,681) - Google's FIRST semantic search invention. Describes DIPRE (Dual Iterative Pattern Relation Expansion) for extracting structured data from unstructured web content. Foundation of the Knowledge Graph.
- Semantic Search (Original) (US 20090070312A1) - Focuses on the meaning and context behind search queries rather than just keywords.
- Phrase-Based Indexing in an Information Retrieval System (US 7,536,408) - Core patent for understanding semantic relationships through phrase co-occurrence. Documents indexed by meaningful phrases that predict other phrases.
- Phrase-Based Detection of Duplicate Documents (US 7,711,679) - Revolutionary approach using phrases rather than individual words to detect and eliminate duplicate documents via related phrase hashes.
- Information Retrieval System for Archiving Multiple Document Versions (US 7,702,618) - Revolutionary approach to indexing and retrieving documents that preserves their evolution over time via MD5 date-range hashes.
- Detecting Spam Documents in a Phrase-Based Information Retrieval System (US 7,603,345) - An innovative approach to identifying and filtering spam documents in search systems using phrase-based analysis and predictive relationships between related phrases.
- Phrase-Based Searching in an Information Retrieval System (US 7,599,914) - A revolutionary approach to information retrieval that transforms how search engines understand and process human language using phrase-based indexing instead of single words. Documents with more related phrases rank higher than those with fewer. Critical for semantic SEO.
- Phrase-Based Generation of Document Descriptions (US 7,584,175) - An innovative information retrieval system that revolutionizes how search engines understand and process documents through sophisticated phrase-based indexing and analysis.
- Phrase-Based Personalization of Searches (US 7,580,929 B1) - A groundbreaking patent that revolutionized how search engines understand user intent and deliver personalized results through sophisticated phrase analysis and machine learning.
- Phrase Identification in Information Retrieval Systems (US 7,580,921 B2) - An innovative approach to information retrieval that uses phrases - not just individual words - to index, search, and organize documents. This system identifies meaningful phrases that predict the presence of other phrases.
- Multiple Index Based Information Retrieval System (US 7,567,959) - A revolutionary approach to indexing and retrieving information using phrases to organize, search, and describe documents at massive scale.
- Automatic Taxonomy Generation in Search Results Using Phrases (US 7,426,507)
- Index Server Architecture Using Tiered and Sharded Phrase Posting Lists (US 7,693,813)
- Index Updating Using Segment Swapping (US 7,702,614)
- Query Scheduling Using Hierarchical Tiers of Index Servers (US 7,925,655)
- Integrating External Related Phrase Information (US 20090070312A1) - Focuses on the meaning and context behind search queries rather than just keywords.
- Detecting Spam Documents (Updated) (US8078629B2) - A revolutionary approach to identifying and filtering spam documents using phrase-based analysis and evaluating standard deviations of related phrase counts.
- Knowledge Graph Based Search System (US 20120158633A1) - Foundation for how entities are analyzed and knowledge graphs are developed.
- Systems and Methods for Ranking Content Using Entity-Based Metrics (US 20180046834A1 (2018)) - Direct foundation for entity salience scoring and EntityAnnotations.
- Identifying Salient Entities in Text (US 20150127617A1 (2015)) - Core patent defining "entity salience" mathematically using dependency graphs.
- Entity Recognition/Identification Model Training (US 9,251,141) - Training models to recognize entities from sentences in knowledge bases.
- Contextual Search Based on Entity Relationships (US 20150370804A1 (2015)) - Uses entity distance, frequency, and co-occurrence for context.
- Query Categorization Based on Entities (US 10169351B1 (2019))
- Ranking Search Results Based on Entity Centrality (US 20160232149A1 (2016))
- Word Sense Disambiguation Using Entity Graphs (US US11687724)
- Generating Structured Information from Unstructured Text (US 20140143273A1 (2014))
- Enterprise Knowledge Graphs Using User-Based Mining (US 20220019908A1)
- Search in Knowledge Graphs (US 12254033B2)
- Searching Quotes of Entities
- Identifying Entity Attribute Relations
- Evaluating an Interpretation for a Search Query (US 20230334045A1) - References BERT for query interpretation accuracy.
- Refining Search Queries
- Query-Dependent Ranking Factors (US 9218397B1)
- Processing and Editing Natural Language Queries
- Natural Language Search Results for Intent Queries (EP3005168A1) - Mechanism for featured snippet eligibility.
- Contextual Estimation of Link Information Gain (WO2020081082) - Information gain scores for ranking - pages with unique information rank higher.
- Combining Parameters of Multiple Search Queries (US 11762848B2)
- Query Composition System (US 20230244657A1)
- High Quality Sites Patent (Panda) (US 9,135,307) - Replaces low-quality search results with high-quality alternatives. Cross-listed with Navneet Panda, April Lehman, and Trystan Upstill inventor sections.
- Predicting Site Quality (N-gram Analysis) (US 9,767,157) - Uses n-grams to evaluate content quality.
- Content Clustering (US 10108694) - Groups websites by topic creating expert clusters for topical authority.
- Context Scoring Adjustments for Answer Passages (US 9,959,315) - Adjusts the relevance score of candidate answer passages based on the surrounding document context to reject misleading short matches.
- Topicality Scores, Social Scores and User-Generated Content
- PageRank (Original) (US 6,285,999B1)
- Reasonable Surfer Model (US 8,117,209B1 / US 7,716,225) - Refines PageRank by considering likelihood of users clicking specific links based on anchor text, font size, position, etc.
- Producing a Ranking for Pages Using Distances in a Web-Link Graph (US 9165040B1)
- Agent Rank (US 7,565,358B2) - Assigns reputation scores to individual authors (agents) based on content quality. The structural ancestor of E-A-T author signals and the layered authority lineage that connects to Procopio (US 8,458,196, per-author per-topic authority) and Overwijk (US 20210004416, site-level topical authority).
- Determining a Quality Measure for a Resource (US 9,558,233B1) - Link value based on actual traffic/clicks. Cross-listed with Paul Haahr (canonical article in his section) and Hyung-Jin Kim's inventor sections.
- Ranking Documents Based on User Behavior and/or Feature Data (US 7,716,225B1) - Refines PageRank by considering likelihood of users clicking specific links based on anchor text, font size, position, etc.
- Document Scoring Based on Document Inception Date (US 8,521,749B2) - Page age and average link rate as ranking factors. Co-invented by Jeff Dean and Paul Haahr; cross-listed with both inventor sections.
- Historical Data Patent (US 7,346,839B2) - Foundational historical-data ranking patent: document aging, freshness, behavioral history and link history.
- Watch Time Based Ranking (US 9,098,511)
- Website Duration Performance (US 9,514,194) - Continuation of the website duration performance family.
- Modifying Search Result Ranking Based on Implicit User Feedback (US 8,661,029B1) - Weighted click-through rate for rankings.
- Click-Through Rate as a Ranking Factor (US 10,229,166B1) - Continuation in the Navboost family. The one cited most in DOJ-trial coverage.
- User-Context-Based Search Engine (US 9,449,105B1) - Evolution from keywords to phrases to context words.
- Personalized Search (US 8,762,373B1)
- Search Result Filters from Resource Content (US 11797626B2)
- Providing Search Results Based on Compositional Query (US 11762933B2) - Entity types and relationships within queries using knowledge graph.
- Contextualizing Knowledge Panels (US 11720577B2)
- Document Activity Logs for Machine Learning (US 20230267277A1)
- Multi-Source Extraction and Scoring of Short Query Answers (US 20230342411A1)
- Presenting Search Result Information (Web Notebooks) (US 11775535B2)
- BERT Question-Answering
- Reader Apparatus for Upconverting Nanoparticle Ink Printed Images - Anti-counterfeiting reader apparatus for upconversion-response nanoparticle inks, providing invisible-taggant document and product security verification. Imported from the Gamma research set as an adjacent-ecosystem patent.
- Digital Aging System - Data-processing technology that applies natural lifecycle concepts (birth, aging, death) to digital information, supporting privacy-by-decay and time-limited data semantics. Imported from the Gamma research set.
- Effects Application Based on Object Clustering (US 2015/0370804 A1) - Apple patent for intelligent media content organization and automated presentation generation through clustering analysis of media objects. Imported from the Gamma research set.
- Smart Optical Input/Output Extension - Mobile data-capture system that embeds camera/OCR functionality into the virtual keyboard interface, enabling context-aware document scanning for workflows. Imported from the Gamma research set.
- Enhanced Information Search System - Database search system combining keyword intelligence with attribute-based filtering. Filed by Nintendo / HAL Laboratory for game-database search. Imported from the Gamma research set as an adjacent-domain example.
Steven Baker, Google Search Patents
- Determining query term synonyms within query context (US 7,636,714) - Foundational patent for context-aware synonym detection. Identifies synonyms by analyzing how candidate replacement terms behave in the same surrounding query context, rather than relying on static thesauri.
- Identifying a synonym with n-gram agreement for a query phrase (US 7,925,498) - Uses n-gram co-occurrence statistics to verify that a candidate synonym preserves the meaning of a multi-word query phrase before substitution.
- Identifying a synonym with N-gram agreement for a query phrase (continuation) (US 8,321,201) - Continuation of the n-gram synonym agreement family, extending coverage to additional phrase-shape variations.
- Document-based synonym generation (US 7,890,521) - Derives synonyms by mining how terms substitute for one another across large web document collections, weighted by document-level co-occurrence signals.
- Document-based synonym generation (continuation) (US 8,161,041) - Continuation refining the document-mined synonym scoring with additional filters against false-positive substitutions.
- Document-based synonym generation (continuation) (US 8,392,413) - Continuation adding refinements to the document-grounded synonym mining pipeline.
- Document-based synonym generation (continuation) (US 8,762,370) - Continuation broadening the document-mined synonym approach to additional content types and contexts.
- Using geographic data to identify correlated geographic synonyms (US 8,041,730) - Detects geographic synonyms (e.g., "NYC" ↔ "New York City") by correlating co-occurring place mentions with shared location signals across documents.
- Using geographic data to identify correlated geographic synonyms (2012 continuation) (US 8,326,866) - Continuation of the geographic synonym detection family, refining correlation signals.
- Using geographic data to identify correlated geographic synonyms (US8417721) (US 8,417,721) - Continuation in the geographic synonyms family with revised claims around correlation thresholds.
- Using geographic data to identify correlated geographic synonyms (US8484188) (US 8,484,188) - Continuation in the geographic synonyms family extending the detection pipeline.
- Using geographic data to identify correlated geographic synonyms (US8527538) (US 8,527,538) - Continuation in the geographic synonyms family covering additional substitution scenarios.
- Abbreviation detection for common synonym generation (US 8,122,022) - Identifies abbreviations and their expansions as synonym pairs to handle short-form query variants (e.g., "info" ↔ "information").
- Longest-common-subsequence detection for common synonyms (US 8,001,136) - Uses longest-common-subsequence string analysis to detect morphological synonym candidates and reject false matches.
- Method and apparatus for generating lexical synonyms for query terms (US 9,183,297) - Generates lexical (form-level) synonyms for query terms by combining linguistic rules with usage statistics.
- Ensuring that a synonym for a query phrase does not drop information present in the query phrase (US 8,661,012) - Verifies that a proposed synonym substitution preserves the informational content of the original query phrase before applying it.
- Identifying related terms in different languages (US 8,798,988) - Cross-lingual related-term detection - maps semantically related terms across languages to support multilingual search expansion.
- System and method for providing search query refinements (US 8,086,619) - Foundational query refinement patent. Suggests refined or expanded versions of a user query based on session, log, and document signals.
- Systems and methods for providing search query refinements (2011 continuation) (US 8,065,316) - Continuation of the search query refinement family.
- Systems and methods for providing search query refinements (2013 continuation) (US 8,504,584) - Continuation in the query refinement family.
- System and method for providing search query refinements (2014 continuation) (US 8,645,407) - Continuation in the query refinement family.
- Systems and methods for providing search query refinements (2016 continuation) (US 9,495,443) - Continuation in the query refinement family.
- System and method for providing search query refinements (2017 continuation) (US 9,552,388) - Continuation in the query refinement family.
- Systems and methods for providing search query refinements (2019 continuation) (US 10,223,439) - Most recent continuation of the query refinement family.
- Locating meaningful stopwords or stop-phrases in keyword-based retrieval systems (US 7,409,383) - Foundational stopword detection patent. Identifies which stopword instances are meaningful in a query (e.g., "the" in "The Who") rather than safely droppable.
- Locating meaningful stopwords (2011 continuation) (US 7,945,579) - Continuation of the meaningful stopword detection family.
- Locating meaningful stopwords (2012 continuation) (US 8,214,385) - Continuation of the meaningful stopword detection family.
- Locating meaningful stopwords (2013 continuation) (US 8,473,510) - Continuation of the meaningful stopword detection family.
- Locating meaningful stopwords (2014 continuation) (US 8,626,787) - Continuation of the meaningful stopword detection family.
- Locating meaningful stopwords (2015 continuation) (US 8,965,919) - Continuation of the meaningful stopword detection family.
- Locating meaningful stopwords (2017 continuation) (US 9,817,920) - Continuation of the meaningful stopword detection family.
- Locating meaningful stopwords (2019 continuation) (US 10,452,718) - Most recent continuation of the meaningful stopword detection family.
- Scoring candidate answer passages (US 9,940,367) - Scores candidate passages from indexed documents as potential answers to a question query, weighing query-match quality, source authority and passage coherence. Core mechanism behind featured snippets and the grounding of generative summaries.
- Scoring candidate answer passages (2020 continuation) (US 10,783,156) - Continuation of the answer passage scoring family.
- Context scoring adjustments for answer passages (US 9,959,315) - Adjusts the relevance score of candidate answer passages based on the surrounding document context to reject misleading short matches.
- Context scoring adjustments for answer passages (2022 continuation) (US 11,409,748) - Continuation of the answer passage context-scoring family.
- Implicit question query identification (US 9,898,554) - Detects whether a non-question-form query is actually expressing an information-seeking question, so it can be routed to question-answering paths.
- Images for query answers (US 10,691,746) - Selects images to accompany or substitute for textual answer passages on question queries.
- Query generation using structural similarity between documents (US 8,346,792) - Generates additional queries by comparing the structural patterns of documents the query already matches, expanding recall while preserving topic.
- Query generation using structural similarity between documents (2015 continuation) (US 9,092,479) - Continuation of the structural-similarity query generation family.
- Query generation using structural similarity between documents (2016 continuation) (US 9,436,747) - Continuation of the structural-similarity query generation family.
- Identifying common co-occurring elements in lists (US 8,037,086) - Identifies semantically related items by analyzing what other elements frequently co-occur with them in user-created lists across the web.
- Identifying common co-occurring elements in lists (2012 continuation) (US 8,285,738) - Continuation of the list co-occurrence family.
- Identifying common co-occurring elements in lists (2013 continuation) (US 8,463,782) - Continuation of the list co-occurrence family.
- Identifying common co-occurring elements in lists (2016 continuation) (US 9,239,823) - Continuation of the list co-occurrence family.
- Golden embeddings (US 11,294,974) - Generates n-dimensional vector embeddings of user interests, web documents, and entities to power similarity-based personalized search and content feeds.
- Validating interests for a search and feed service (US 11,200,288) - Validates inferred user interests before feeding them into the search-and-feed personalization pipeline, reducing spurious topic injections.
- User interest modeling (US App. 15/445,927) - Published application - models user interest signals to drive personalized search and content surfacing.
- Enhanced search for generating a content feed (US App. 15/445,931) - Published application - extends search machinery to build a personalized content feed surface as an output of search retrieval.
Amit Singhal, Google Search Patents
- Interleaving search results (US 8,086,600) - Foundational interleaving patent. Combines results from multiple ranking algorithms or scoring strategies into a single result list by alternating between them, enabling A/B-style comparison of ranking models in live serving.
- Interleaving search results (2014 continuation) (US 8,738,597) - Continuation refining the interleaving mechanics for production ranking experiments.
- Interleaving search results (2015 continuation) (US 9,002,817) - Continuation extending interleaved evaluation to additional ranking scenarios.
- Merging search results (US 8,392,394) - Merges results from multiple sub-engines or verticals (web, news, image, video) into a single unified result page, preserving rank quality across heterogeneous sources.
- Systems and methods for correlating document topicality and popularity (US 8,595,225) - Combines topical relevance signals with popularity signals to score documents, calibrating how much weight popularity gets per query type to avoid drowning niche topical matches under popular but less relevant pages.
- Ranking nodes in a linked database based on node independence (US 8,719,276) - A PageRank-style extension that scores nodes (documents) by how independent their inbound link sources are, downweighting clusters of mutually reinforcing links to resist link-farm manipulation.
- Methods and systems for efficient query rewriting (US 7,840,547) - Foundational query rewriting patent. Rewrites a user query into an alternative form that retrieves better results, balancing fidelity to the original intent with broader coverage of the relevant document set.
- Methods and systems for efficient query rewriting (2014 continuation) (US 8,631,026) - Continuation refining the rewrite scoring and fallback behavior.
- Efficient query rewriting (2015 continuation) (US 9,165,033) - Continuation in the query rewriting family.
- Methods and systems for efficient query rewriting (2020 continuation) (US 10,685,017) - Most recent continuation of the query rewriting family.
- Search queries improved based on query semantic information (US 8,055,669) - Foundational semantic query improvement patent. Augments queries with semantic information (entities, concepts, intent labels) extracted from query analysis to drive better retrieval matching beyond literal keyword overlap.
- Search queries improved based on query semantic information (2013 continuation) (US 8,577,907) - Continuation of the semantic query improvement family.
- Determining quality of linked documents (US 7,783,639) - Foundational link quality patent. Evaluates the quality of a document partly by analyzing the documents that link to it, going beyond raw inbound-link count to include link-source authority and contextual relevance.
- Determining quality of linked documents (2012 continuation) (US 8,176,056) - Continuation refining the link-source authority scoring for document quality.
- Determining quality of linked documents (2014 continuation) (US 8,825,645) - Continuation of the link quality family.
- Methods and systems for improving a search ranking using location awareness (US 7,606,798) - Foundational location-aware ranking patent. Uses the geographic location associated with a query or user to adjust the ranking of documents, surfacing local results when location intent is detected.
- Ranking documents based on a location sensitivity factor (US 8,171,048) - Continuation introducing an explicit location sensitivity factor that decides per-query how much weight to give location-based ranking signals.
- System and method for providing preferred language ordering of search results (US 7,451,129) - Ranks search results by language preference inferred from the user, the query, or session signals, so results in the user's preferred language surface above otherwise-equivalent results in other languages.
- System and method for providing preferred country biasing of search results (US 7,451,130) - Applies a country preference bias to ranking so that results from the user's country (or a query-inferred country of interest) get a configurable score boost.
- Systems and methods for detecting commercial queries (US 8,046,350) - Foundational commercial-query detection patent. Classifies whether a search query expresses commercial intent (transactional or comparison shopping) so the engine can adjust result mix, ad inventory, and result-page format accordingly.
- Systems and methods for detecting commercial queries (2013 continuation) (US 8,510,289) - Continuation refining the commercial-query classifier.
- Serving advertisements using user request information and user information (US 8,352,499) - Combines per-request signals (current query, recent context) with per-user signals (profile, history) to pick advertisements to serve, tightening relevance over either signal alone.
Prabhakar Raghavan, Google Search Patents
- User-sensitive PageRank (US 7,624,104) - Personalizes the PageRank computation per user by biasing the random-walk teleportation vector toward documents the user has visited or rated, producing per-user authority scores rather than a single global ranking.
- User-sensitive PageRank (2016 continuation) (US 9,495,452) - Continuation of the user-sensitive PageRank family with refinements to the teleportation biasing.
- System and method for ranking hyperlinked documents based on a stochastic backoff process (US 6,792,419) - Extends PageRank with a stochastic backoff model: when a user does not find satisfactory content at the current node, they "back off" probabilistically through the link graph. The model captures realistic user navigation and produces more robust authority scores than uniform random walks.
- Method and system for trawling the World-wide Web to identify implicitly-defined communities of web pages (US 6,886,129) - Discovers implicit topical communities on the web by detecting dense bipartite subgraphs of pages that link to overlapping sets of resources. Foundational work in graph-based community detection on the web that predates modern entity clustering.
- Apparatus and method for adaptively ranking search results (US 6,738,764) - Ranks search results adaptively based on user feedback (clicks, dwell, explicit ratings) so the ranking system improves over time per query, per user, or per query class. Earlier formalization of click-as-implicit-relevance that underlies modern learning-to-rank pipelines.
- Method and apparatus for merging result lists from multiple search engines (US 6,728,704) - Combines ranked result lists from multiple search engines (or multiple ranking algorithms in one engine) into a single merged list, normalizing scores across sources and resolving overlap. Foundational meta-search technique still used in universal-search merging today.
- Multilevel taxonomy based on features derived from training documents classification using fisher values as discrimination values (US 6,233,575) - Builds multilevel hierarchical document taxonomies using Fisher discrimination values to select features at each level of the hierarchy, producing classification trees that are far more discriminating than flat or single-level approaches. Used as a building block in topic classification systems across search engines.
- Automatic subspace clustering of high dimensional data for data mining applications (US 6,003,029) - Algorithm for finding clusters in subspaces of high-dimensional data (CLIQUE algorithm). Foundational data-mining work that addresses the curse of dimensionality by searching for clusters in lower-dimensional projections rather than the full feature space.
- Method and system for filtering of information entities (US 6,996,572) - Filters information entities (documents, web pages, news items) against user-specified or learned interest profiles, removing items predicted to be irrelevant before they reach the user. Early work in personalized content filtering that underlies feed personalization today.
- Method for interactively creating an information database including preferred information elements, such as, preferred-authority, world wide web pages (US 6,336,112) - Interactive system for building a curated information database where users designate preferred-authority sources for each topic. Predates and informs modern knowledge-graph editorial curation and user-personalized topical authority.
- Methods and systems for providing a customizable guide for navigating a corpus of content (US 8,914,729) - Generates customizable navigation guides through a content corpus so users can traverse related items along multiple axes (topic, date, popularity, user-defined paths) rather than being limited to a single linear ranking.
- System and method for obtaining of user generated content boxes (US 8,856,097) - Surfaces user-generated content (reviews, comments, social posts) inside search result pages as dedicated content boxes, integrating community-authored material into the ranking surface alongside traditional documents.
- Determining information about a location based on travel related to the location (US 9,251,168) - Derives quality and relevance signals for a location by analyzing travel patterns (how far people travel to reach it, return rate, time spent). Travel investment becomes a behavioral quality signal beyond passive popularity counts.
- Determining the quality of locations based on travel time investment (US 9,558,210) - Continuation refining the travel-time quality signal with explicit time-investment scoring.
- System and method for visualizing the temporal evolution of object metadata (US 7,581,184) - Tracks and visualizes how metadata on documents and entities changes over time, surfacing temporal patterns (trending topics, fading interest, recurring cycles) that static metadata snapshots miss.
- System and method for selecting object metadata evolving over time (US 7,739,275) - Continuation of the temporal metadata family, focused on selection of the metadata snapshot most relevant to a given query time.
Nitin Gupta, Google Search Patents
- Query suggestion templates (US 9,529,856) - Generates query suggestion templates from observed query patterns. Templates capture the shape of common query intents and can be filled in to produce concrete suggestions tailored to a current query.
- Query suggestion templates (2020 continuation) (US 10,635,717) - Continuation of the query suggestion templates family.
- Query Suggestion Templates (application) (US App. 20,140,358,940) - Published application of the query suggestion templates invention.
- Query Suggestion Templates (2017 application) (US App. 20,170,039,209) - Application continuation of the query suggestion templates family.
- Query suggestions based on entity collections of one or more past queries (US 9,342,626) - Generates query suggestions from entity collections derived from a user's past queries. The entity collection acts as a topical anchor that shapes the suggestions the system offers.
- Query suggestions based on entity collections (2019 continuation) (US 10,360,225) - Continuation of the entity-based query suggestions family.
- Query suggestions based on entity collections (2023 continuation) (US 11,853,307) - Continuation of the entity-based query suggestions family.
- Query suggestions based on entity collections (2025 continuation) (US 12,216,669) - Continuation of the entity-based query suggestions family.
- Query Suggestions Based On Entity Collections (2025 application) (US App. 20,250,258,828) - Most recent published application of the entity-based query suggestions family.
- Query rewrites for generating auto-complete suggestions (US 9,235,654) - Generates autocomplete suggestions by rewriting the partial user query into more complete alternatives, expanding the suggestion surface beyond literal prefix matches.
- Populating query suggestion database using chains of related search queries (US 9,342,600) - Builds the query suggestion database from chains of related search queries observed across sessions, mining the natural sequence in which users refine queries to construct the suggestion corpus.
- Semantically equivalent query templates (US 10,073,882) - Identifies query templates that are semantically equivalent (express the same intent in different surface forms) and groups them together so suggestion and routing logic can treat them as one underlying need.
- Reranking query completions (US 9,298,852) - Reranks query completion candidates produced by the autocomplete system, using contextual signals beyond raw frequency to surface the completions most likely to match the user's current intent.
- Reranking Query Completions (application) (US App. 20,150,169,578) - Published application of the query completion reranking invention.
- Personalized suggestions based on past queries (US 10,496,649) - Personalizes query suggestions by drawing on the individual user's history of past queries, producing suggestions tuned to the user's recurring topical interests rather than population averages.
- Supplementing search results with historically selected search results of related queries (US 9,298,828) - Adds historically selected (clicked, dwelled on) results from related past queries to the current result set, enriching retrieval with documents the engine knows have satisfied similar intents before.
- Generating elements of answer-seeking queries and elements of answers (US 10,592,540) - Decomposes answer-seeking queries into elemental components (entity, attribute, qualifier) and constructs answer elements that align element-to-element with the query, supporting structured direct-answer generation.
- Generating Elements of Answer-Seeking Queries (application) (US App. 20,170,011,116) - Application of the answer-element decomposition invention.
- Answer to question neural networks (US 11,093,813) - Neural-network architecture for generating an answer to a question query. Bridges the symbolic answer-passage scoring era to learned end-to-end answer generation.
- Answer to Question Neural Networks (application) (US App. 20,180,114,108) - Published application of the answer-question neural network invention.
- Candidate answer passages (US 10,180,964) - Extracts and scores candidate answer passages from a corpus to identify the passage most likely to answer the user's question. Gupta's contribution to the answer-passage family.
- Controlled content diversity in retrieval for generative search (US App. 20,260,072,965) - Most recent in the set. Controls content diversity in the retrieval stage that feeds generative search (AI Overviews / SGE). Ensures the generative model receives a diverse, non-redundant set of source passages so the synthesized answer is comprehensive and balanced.
- Content knowledge query generation through computer analysis (US App. 20,230,368,693) - Generates queries from analyzed content knowledge to drive retrieval and answer generation, automating what previously required hand-authored query templates.
- Data facet generation and recommendation (US App. 20,230,401,457) - Generates facets (dimensions) over a data set and recommends the facets most useful to surface for a given query, supporting faceted-search and explore-style result presentation.
- Card interface for managing domain search results (US App. 20,150,212,710) - Presents domain-search results as cards in an interface that lets users manage and pivot across the result set, predating modern entity-card SERP modules.
- Search Engine Optimization of Domain Names and Websites (US App. 20,160,043,989) - Direct SEO patent. Describes mechanisms for optimizing domain names and websites for search-engine retrieval, including signals the system reads from domain-level metadata and on-page structure.
- System and method for identifying search results satisfying a search query (US 9,405,834) - Identifies which retrieved documents actually satisfy a search query, distinguishing topical match from genuine satisfaction of the user's information need.
- Query generation using structural similarity between documents (US 8,346,792) - Co-invented with Steven Baker. The full article is in the Baker section.
- Query generation using structural similarity between documents (2015 continuation) (US 9,092,479) - Co-invented continuation. The full article is in the Baker section.
- Query generation using structural similarity between documents (2016 continuation) (US 9,436,747) - Co-invented continuation. The full article is in the Baker section.
- Implicit question query identification (US 9,898,554) - Co-invented with Steven Baker. The full article is in the Baker section.
- Context scoring adjustments for answer passages (US 9,959,315) - Co-invented with Steven Baker. The full article is in the Baker section.
Navneet Panda, Google Search Patents
- Ranking search results (US 8,682,892) - The foundational Panda patent. Ranks search results by combining standard relevance with a per-site quality score derived from site-level signals, demoting content from low-quality sites even when their pages match the query well.
- Ranking search results (2017 continuation) (US 9,684,697) - Continuation of the foundational Panda ranking family.
- Ranking search results (2018 continuation) (US 10,055,467) - Most recent continuation of the Panda ranking family.
- Site quality score (US 9,031,929) - Computes a per-site quality score from multiple signals (content depth, link quality, user-behavior, structural integrity) that the ranking system consumes as a site-level multiplier on per-page relevance.
- Site quality score (2017 continuation) (US 9,760,641) - Continuation of the site quality score family.
- Predicting site quality (US 9,767,157) - Predicts the quality of a site from signals available before the site has accumulated enough behavioral data, enabling the ranking system to apply a quality estimate to new sites rather than waiting for full ground-truth quality signals to develop.
- Predicting Site Quality (application) (US App. 20,140,280,011) - Application form of the predicting site quality invention.
- Website duration performance based on category durations (US 9,171,086) - Derives a site quality signal from how long users spend on the site relative to category-typical durations. Sites whose dwell durations underperform category norms get a quality penalty.
- Website duration performance (2016 continuation) (US 9,514,194) - Continuation of the website duration performance family.
- Locally significant search queries (US 9,348,925) - Identifies search queries that are significant in a specific geographic locality but not at global scale, supporting hyperlocal ranking adjustments and locally tuned query suggestions.
- Locally Significant Search Queries (application) (US App. 20,140,172,843) - Application form of the locally significant queries invention.
- Selectively generating alternative queries (US 9,135,307) - Decides whether to generate alternative queries for a current query based on configured selection criteria, avoiding unnecessary rewrites when the original query is already well-formed and likely to satisfy intent.
- Revising search queries (US 9,449,095) - Revises an incoming search query to improve retrieval, applying configured revision rules that handle spelling, abbreviation, intent broadening, and topical refinement.
- Learning concept templates from web images to query personal image databases (US 8,958,661) - Learns concept templates from web images so users can query their personal image databases using natural-language concepts (e.g., "beach photos", "birthday parties") rather than requiring per-image manual tagging.
- Learning Concept Templates From Web Images (application) (US App. 20,080,240,575) - Application form of the image concept template invention.
Anand Shukla, Google Search Patents
- Search with stateful chat (US App. 20,240,289,407) - 2024 patent bridging classical search to LLM-based conversational search. Maintains conversation state across multiple turns so each new query is interpreted in the context of the ongoing chat session, blending retrieval and generation in a single user-facing dialog surface.
- Surfacing in-depth articles in search results (US 9,996,624) - Identifies in-depth long-form articles for a query and surfaces them in search results in a dedicated module, distinguishing them from news, quick-answer, and standard web results. Powers the original "In-Depth Articles" SERP feature.
- Surfacing In-Depth Articles (application) (US App. 20,150,379,140) - Published application of the in-depth article surfacing invention.
- Content channel curation (US App. 20,190,266,283) - Curates content channels by combining editorial signals with user behavior, producing channel collections that are coherent topically while reflecting the audience's evolving interests.
- Query pattern matching (US 11,023,506) - Matches incoming queries against learned query patterns to identify intent and template fillers. Modern pattern-matching approach that handles paraphrasing and word-order variation.
- Query Pattern Matching (application) (US App. 20,200,133,967) - Published application of the query pattern matching invention.
- Query topic map (US App. 20,190,266,288) - Maps queries into a topical taxonomy to drive retrieval, recommendations, and topical routing. The query topic map is the substrate for topic-aware ranking adjustments.
- Query augmentation (US 9,128,945) - Augments incoming queries with additional terms derived from contextual signals to improve retrieval recall while preserving intent.
- Query augmentation (2018 continuation) (US 9,916,366) - Continuation of the query augmentation family.
- Search augmentation (US 8,346,791) - Earlier search augmentation work. Improves retrieval by augmenting the search process with auxiliary signals beyond the original query terms.
- Generate an index for enhanced search based on user interests (US App. 20,180,246,899) - Generates a specialized index optimized for enhanced search retrieval keyed to user interests. Companion to the user-interest-modeling family in the Baker section; this canonical describes the indexing variant.
- Enhanced search to generate a feed based on user interests (2024 grant) (US 12,032,638) - 2024 granted continuation of the enhanced search feed family.
- Enhanced Search To Generate A Feed (2024 continuation) (US App. 20,240,220,554) - Most recent published application in the enhanced-search-feed family.
- Golden embeddings (US 11,294,974) - Co-invented with Steven Baker. The full article is in the Baker section.
- User interest modeling (US App. 20,180,246,973) - Co-invented with Steven Baker. The full article is in the Baker section.
- Enhanced search for generating a content feed (US App. 20,180,246,974) - Co-invented with Steven Baker. The full article is in the Baker section.
Krishna Bharat, Google Search Patents
- Ranking Search Results by Reranking Based on Local Inter-Connectivity (Hilltop Algorithm) (US 6,526,440) - The Hilltop algorithm. Identifies "expert documents" on a topic, then ranks results by the inter-connectivity among experts who reference the candidate, distinguishing genuinely authoritative pages from heavily-linked but non-authoritative ones.
- Ranking Search Results by Reranking Based on Local Inter-Connectivity (continuation) (US 6,725,259) - Continuation of the Hilltop algorithm patent. Refines the expert-document identification and inter-connectivity scoring.
- System and Method for Supporting Editorial Opinion in the Ranking of Search Results (US 7,096,214) - Incorporates editorial opinion (curated boosts and demotions) into algorithmic ranking, blending human curation with mechanical retrieval.
- System and Method for Supporting Editorial Opinion in the Ranking of Search Results (continuation) (US 7,386,543) - Continuation refining how editorial signals combine with algorithmic ranking.
- Identification of Semantic Units from Within a Search Query (US 7,249,121) - Detects multi-word semantic units (named entities, fixed phrases) within a query and treats them as single retrieval atoms, raising precision for queries where word-level decomposition would scatter the intent.
- Identification of Semantic Units from Within a Search Query (continuation) (US 8,321,410) - Continuation refining the multi-word unit detection.
- Identification of Semantic Units from Within a Search Query (later continuation) (US 8,719,262) - Further refinement of semantic-unit identification.
- Query Rewriting with Entity Detection (US 7,536,382) - Detects entities in the query and rewrites the query to match how documents naturally reference those entities, recovering documents that would not match the literal query but do match the entity-aware rewrite.
- Query Rewriting with Entity Detection (continuation 2012) (US 8,112,432) - Continuation of the query rewriting with entity detection family.
- Query Rewriting with Entity Detection (continuation 2013-1) (US 8,452,799) - Continuation of the query rewriting with entity detection family.
- Query Rewriting with Entity Detection (continuation 2013-2) (US 8,805,867) - Continuation of the query rewriting with entity detection family.
- Query Rewriting with Entity Detection (app 2005) (US App 2005/0222977) - Earliest published application for the query rewriting with entity detection family.
- Query Rewriting with Entity Detection (app 2009) (US App 2009/0204592) - Continuation application.
- Query Rewriting with Entity Detection (app 2012) (US App 2012/0136885) - Continuation application.
- Query Rewriting with Entity Detection (app 2013) (US App 2013/0262499) - Continuation application.
- Systems and Methods for Improving the Ranking of News Articles (US 7,577,655) - Ranks news articles by combining article relevance with publisher authority, freshness, and topical cluster signals. Foundational news-search ranking patent.
- Systems and Methods for Improving the Ranking of News Articles (continuation 2012-1) (US 8,126,876) - Continuation in the news-ranking family.
- Systems and Methods for Improving the Ranking of News Articles (continuation 2012-2) (US 8,332,382) - Continuation in the news-ranking family.
- Systems and Methods for Improving the Ranking of News Articles (continuation 2014) (US 8,645,368) - Continuation in the news-ranking family.
- Systems and Methods for Improving the Ranking of News Articles (app 2005) (US App 2005/0060312) - Earliest application in the news-ranking family.
- Systems and Methods for Improving the Ranking of News Articles (app 2009) (US App 2009/0276429) - Continuation application.
- Systems and Methods for Improving the Ranking of News Articles (app 2012) (US App 2012/0158711) - Continuation application.
- Systems and Methods for Improving the Ranking of News Articles (app 2013) (US App 2013/0159294) - Continuation application.
- Methods and Apparatus for Employing Usage Statistics in Document Retrieval (US 8,001,118) - Uses aggregate usage statistics (clicks, dwell, repeated visits) as ranking inputs, raising documents that demonstrate sustained engagement above documents that merely match the query text.
- Methods and Apparatus for Employing Usage Statistics in Document Retrieval (continuation 2012-1) (US 8,156,100) - Continuation refining usage-statistics-based retrieval.
- Methods and Apparatus for Employing Usage Statistics in Document Retrieval (continuation 2013) (US 8,352,452) - Continuation refining usage-statistics-based retrieval.
- Methods and Apparatus for Employing Usage Statistics in Document Retrieval (app 2002) (US App 2002/0123988) - Earliest published application for the usage-statistics family.
- Methods and Apparatus for Employing Usage Statistics in Document Retrieval (app 2011) (US App 2011/0179023) - Continuation application.
- Methods and Apparatus for Employing Usage Statistics in Document Retrieval (app 2012) (US App 2012/0226705) - Continuation application.
- Methods and Apparatus for Ranking Documents (US 8,090,717) - Generalized document ranking framework combining multiple signal sources (link analysis, content matching, usage data) into a unified score, with per-signal weighting tuned for different query types.
- Methods and Apparatus for Ranking Documents (continuation 2014) (US 8,843,479) - Continuation refining the multi-signal ranking framework.
- Methods and Apparatus for Ranking Documents (continuation 2016) (US 9,477,714) - Continuation in the document-ranking family.
- Methods and Apparatus for Ranking Documents (continuation 2019) (US 10,496,652) - Latest continuation in the document-ranking family.
- Query Augmentation (US 9,128,945) - Augments user queries with additional terms or constraints derived from query context, user profile, and historical reformulation patterns, increasing recall while preserving precision.
- Query Augmentation (continuation 2018) (US 9,916,366) - Continuation refining query augmentation methods.
- Search Augmentation (US 8,346,791) - Augments search results with related entities, refinement suggestions, and contextual information panels derived from the result set and the query intent.
- Image Selection for News Search (US 8,775,436) - Selects representative images for news search results by combining image-content quality signals with relevance to the article and the query, ensuring carousel and panel imagery is visually appropriate.
- Image Selection for News Search (continuation) (US 9,613,061) - Continuation refining image selection for news.
- Method and Apparatus for Query-Specific Bookmarking and Data Collection (US App 2005/0114299) - Bookmarks results and accompanying notes in the context of the originating query, so users can return to result sets organized by intent rather than by raw URL list.
- Method for Estimating Coverage of Web Search Engines (US App 2005/0055342) - Statistical method for estimating how much of the web a search engine has indexed, using cross-engine query sampling and overlap analysis.
- Using an Expanded View to Display Links Related to a Topic (US 9,678,618) - Expands a search result or content panel to display additional links curated by topical relationship to the focal item, surfacing related entities and adjacent topics in an expanded-view UI.
- Entity Identification Model Training (US 9,251,141) - Trains entity-recognition models using complete sentences from authoritative sources, learning to predict entities even from fragmentary text. Cross-listed with the 65 Google Patents collection as pat-21.
- Rendering Advertisements with Documents Having One or More Topics Using User Topic Interest (US 7,346,606) - Renders contextual advertisements alongside documents by matching document topics to user topic-interest profiles, enabling topic-aware ad selection rather than pure keyword matching.
- Rendering Advertisements with Documents (continuation 2012-1) (US 8,090,706) - Continuation in the topic-interest ads family.
- Rendering Advertisements with Documents (continuation 2012-2) (US 8,296,285) - Continuation in the topic-interest ads family.
- Rendering Advertisements with Documents (app 2004) (US App 2004/0267723) - Earliest published application for the topic-interest ads family.
- Rendering Advertisements with Documents (app 2008) (US App 2008/0097833) - Continuation application.
- Rendering Advertisements with Documents (app 2012) (US App 2012/0072291) - Continuation application.
Srinivasan Venkatachary, Google Search Patents
- Generative Summaries for Search Results (US 11,769,017) - Canonical patent for the generative-summaries pipeline that powers AI Overviews and Search Generative Experience. Combines retrieval of grounded passages with an LLM that composes a synthesized answer attributable to the underlying sources.
- Generative Summaries for Search Results (continuation 2024) (US 11,886,828) - Continuation in the generative-summaries family.
- Generative Summaries for Search Results (continuation 2024b) (US 11,900,068) - Continuation refining the generative-summaries pipeline.
- Generative Summaries for Search Results (continuation 2024c) (US 12,118,325) - Later continuation in the generative-summaries family.
- Generative Summaries for Search Results (app 2024) (US App 2024/0220735) - Continuation application in the generative-summaries family.
- Generative Summaries for Search Results (app 2025) (US App 2025/0005303) - Latest application in the generative-summaries family.
- Enhanced Search to Generate a Feed Based on a User's Interests (US 12,032,638) - Generates a personalized content feed by treating each saved user interest as an implicit search query, scoring fresh results against the interest plus user-engagement signals.
- Enhanced Search to Generate a Feed Based on a User's Interests (app 2018) (US App 2018/0246972) - Earliest published application in the enhanced-search-feed family.
- Enhanced Search to Generate a Feed Based on a User's Interests (app 2024) (US App 2024/0220554) - Continuation application in the enhanced-search-feed family.
- Enhanced Search for Generating a Content Feed (US App 2018/0246974) - Companion application in the enhanced-search-feed family covering content-feed generation paths.
- Candidate Answer Passages (US 10,180,964) - Identifies candidate answer passages within retrieved documents. Each candidate carries features for downstream scoring. Foundation for featured-snippet extraction.
- Scoring Candidate Answer Passages (US 9,940,367) - Scores candidate answer passages on query-match quality, source authority, and passage coherence. Selects the best span for surfacing as a featured snippet or grounding for a generative summary.
- Scoring Candidate Answer Passages (continuation) (US 10,783,156) - Continuation refining passage scoring with additional features.
- Context Scoring Adjustments for Answer Passages (US 9,959,315) - Adjusts passage scores using the surrounding context within the source document, so a passage in a topically-coherent section scores higher than the same passage marooned in mixed content. Cross-listed with 65 Google Patents as pat-42.
- Context Scoring Adjustments for Answer Passages (continuation) (US 11,409,748) - Continuation refining context-aware passage scoring.
- Implicit Question Query Identification (US 9,898,554) - Detects when a query that does not appear question-shaped on the surface is implicitly asking a question, so the system can route it to question-answering paths instead of standard retrieval.
- Implicit Question Query Identification (app) (US App 2015/0142851) - Earliest application in the implicit-question-query family.
- Keyword-Based Conversational Searching Using Voice Commands (US 9,305,064) - Conversational voice-search framework that maintains keyword context across multi-turn voice queries, supporting follow-up questions without restating prior context.
- Query Generation Using Structural Similarity Between Documents (US 8,346,792) - Generates expansion queries by analyzing structural similarity between documents (template patterns, section headings, layout), bridging template detection and query rewriting, so retrieval finds structurally-equivalent content the user did not literally search for.
- Query Generation Using Structural Similarity Between Documents (continuation 2015) (US 9,092,479) - Continuation refining structural-similarity-based query generation.
- Query Generation Using Structural Similarity Between Documents (continuation 2016) (US 9,436,747) - Continuation in the structural-similarity family.
- Resource Identification from Organic and Structured Content (US 8,688,714) - Identifies resources (entities, documents, products) referenced within content by analyzing both organic prose and structured-data markup, enabling resource-aware retrieval beyond literal keyword matching.
- Resource Identification from Organic and Structured Content (continuation) (US 8,688,715) - Continuation in the resource-identification family.
- Content Channel Curation (US App 2019/0266283) - Curates content channels by clustering related items around topical themes, supporting topical browsing surfaces like Discover. Cross-listed with the Anand Shukla inventor section.
- Emoji Classifier (US App 2019/0258719) - Classifies emoji usage in text to inform sentiment analysis, content classification, and contextual content delivery, recognizing that emoji carry topical and emotional signal beyond their literal Unicode meaning.
Jeromy William Henry, Google Search Patents
- Providing Knowledge Panels with Search Results (US 9,454,611) - Foundational Knowledge Panel display patent. Defines how the entity panel appears alongside SERP results, what facts it surfaces, and how it interacts with the result list.
- Providing Knowledge Panels with Search Results (app 2016) (US App 2016/0210364) - Earliest published application in the knowledge-panels family.
- Providing Knowledge Panels with Search Results (continuation 2018) (US 10,318,567) - Continuation refining the Knowledge Panel display logic.
- Providing Knowledge Panels with Search Results (app 2019) (US App 2019/0278793) - Continuation application in the knowledge-panels family.
- Providing Knowledge Panels with Search Results (continuation 2021) (US 11,093,539) - Later continuation in the knowledge-panels family.
- Providing Knowledge Panels with Search Results (app 2021) (US App 2021/0374171) - Continuation application in the knowledge-panels family.
- Providing Knowledge Panels with Search Results (continuation 2023) (US 11,836,177) - Continuation refining Knowledge Panel composition.
- Providing Knowledge Panels with Search Results (app 2024) (US App 2024/0202224) - Latest continuation application in the knowledge-panels family.
- Providing Knowledge Panels with Search Results (app 2016b) (US App 2016/0350414) - Companion published application in the knowledge-panels family.
- Structured Entity Information Page (US 10,110,701) - Foundational patent for the standalone entity-detail page. When users click into an entity, the system renders a structured page with facts, related entities, content sources, and interaction surfaces.
- Structured Entity Information Page (app 2016) (US App 2016/0173644) - Earliest published application in the structured-entity-page family.
- Structured Entity Information Page (app 2019) (US App 2019/0082031) - Continuation application in the structured-entity-page family.
- Structured Entity Information Page (continuation 2020) (US 10,587,726) - Continuation refining the structured entity page.
- Structured Entity Information Page (app 2020) (US App 2020/0220947) - Continuation application in the structured-entity-page family.
- Structured Entity Information Page (app 2022) (US App 2022/0086258) - Continuation application in the structured-entity-page family.
- Structured Entity Information Page (continuation 2023) (US 11,706,318) - Continuation in the structured-entity-page family.
- Structured Entity Information Page (app 2023) (US App 2023/0319164) - Continuation application in the structured-entity-page family.
- Structured Entity Information Page (continuation 2024) (US 12,149,599) - Latest grant in the structured-entity-page family.
- Structured Entity Information Page (app 2025) (US App 2025/0231954) - Latest application in the structured-entity-page family.
- Personalized Entity Information Page (US 10,917,499) - Personalization layer over the structured entity page. Content selection and ordering shape to the viewing user, so two users see different facts and emphasis on the same entity.
- Personalized Entity Information Page (app 2021) (US App 2021/0126987) - Continuation application in the personalized-entity-page family.
- Personalized Entity Information Page (continuation 2022) (US 11,218,568) - Continuation in the personalized-entity-page family.
- Search Results with Structured Image Sizes (US 9,373,155) - Image-sizing rules for SERP imagery. Defines when to surface large vs small images, how to optimize for visual density, and how to adapt to device form factor.
- Search Results with Structured Image Sizes (app) (US App 2015/0161764) - Earliest published application in the SERP image-sizing family.
- Presenting Secondary Music Search Result Links (US 9,990,425) - Music-search-specific UI for surfacing related artist, album, and song links alongside the primary music search result, supporting discovery of adjacent musical content.
Ramanathan V. Guha, Google Search Patents
- Programmable Search Engine (US 7,693,830) - Foundational patent for Google Custom Search Engine. Lets third parties define their own search engines by specifying refinements, scoring rules, and domain restrictions over the Google index.
- Programmable Search Engine (continuation 2012) (US 8,316,040) - Continuation in the programmable-search-engine family.
- Programmable Search Engine (continuation 2015) (US 9,031,937) - Continuation refining programmable-search-engine logic.
- Programmable Search Engine (app 2010) (US App 2010/0217756) - Continuation application in the programmable-search-engine family.
- Programmable Search Engine (app 2013) (US App 2013/0124510) - Continuation application.
- Programmable Search Engines (app 2016) (US App 2016/0299983) - Continuation application.
- Aggregating Context Data for Programmable Search Engines (US 7,716,199) - Aggregates context data (user profile, location, prior queries) for programmable search engines so refinements respect rich context. Same family as the canonical programmable-search patent.
- Aggregating Context Data (continuation 2011) (US 8,051,063) - Continuation in the aggregating-context-data family.
- Aggregating Context Data (app 2010) (US App 2010/0250513) - Continuation application in the aggregating-context-data family.
- Aggregating Context Data (continuation 2014) (US 8,756,210) - Continuation refining context aggregation for programmable search.
- Generating Specialized Search Results in Response to Patterned Queries (US 7,593,939) - Pattern-matched query handling for programmable search engines, allowing custom engines to define query templates and produce specialized result formats.
- Dynamic Specification of Custom Search Engines at Query-Time (US 8,892,552) - Lets the search system construct a custom search engine on the fly per query, dynamically applying refinements without requiring a pre-defined CSE.
- Generating and Presenting Advertisements Based on Context Data for Programmable Search Engines (US App 2013/0110627) - Ad-rendering pipeline keyed on the context data programmable search engines collect, enabling context-aware advertising on custom search surfaces.
- Search Result Ranking Based on Trust (US 8,352,467) - Ranks search results using trust signals derived from social and topical neighborhoods, so results from trusted sources rise even when literal relevance is comparable.
- Search Result Ranking Based on Trust (continuation 2014) (US 8,818,995) - Continuation refining trust-based ranking.
- Search Result Ranking Based on Trust (continuation 2019) (US 10,268,641) - Latest continuation in the trust-based ranking family.
- Detecting Spam Related and Biased Contexts for Programmable Search Engines (US 7,743,045) - Detects spam and biased contexts within programmable search engine results, so user-defined CSEs cannot become spam-amplification surfaces.
- Detecting Spam Related and Biased Contexts (app 2010) (US App 2010/0223250) - Continuation application in the spam-detection family.
- Detecting Spam Search Results for Context Processed Search Queries (US 8,452,746) - Continuation in the spam-detection family with broader query-context coverage.
- Classifying Search Results (US 9,043,322) - Classifies search results into structured categories so the SERP layout adapts: shopping results in shopping panels, news in news cards, etc.
- Classifying Search Results to Determine Page Elements (US 8,600,987) - Continuation in the classifying-search-results family with page-element granularity.
- Classifying Search Results to Determine Page Elements (app) (US App 2014/0089305) - Continuation application in the classifying-search-results family.
- Propagating Promotional Information on a Social Network (US 9,466,077) - Propagates promotional content through a social network using graph-walk and engagement signals, so promotions reach interested users via the social graph rather than broadcast targeting.
- Propagating Promotional Information (continuation 2018) (US 10,074,109) - Continuation refining social propagation.
- Propagating Promotional Information (app 2010) (US App 2010/0332330) - Earliest application in the social-propagation family.
- Propagating Promotional Information (app 2016) (US App 2016/0379273) - Continuation application in the social-propagation family.
- Customized Web Summaries and Alerts Based on Custom Search Engines (US 8,725,716) - Generates personalized web summaries and digest alerts by running custom search engines on a schedule and surfacing changes to the user, supporting daily-digest patterns over user-defined topical scopes.
- Customized Web Summaries and Alerts (continuation) (US 9,323,853) - Continuation refining customized web summaries and alerts.
- Filtering Search Results Using Annotations (US 8,341,150) - Filters search results using user-supplied annotations on documents, supporting collaborative curation and personalized result filtering through declarative tags.
- Generating and Displaying Tasks (US App 2014/0172853) - Generates task-shaped result surfaces (book a flight, send a message, schedule a meeting) directly from queries, so search becomes an action launcher rather than just a document retriever.
- Generating and Displaying Tasks (companion app) (US App 2014/0156623) - Companion application in the generating-tasks family.
- Contextual Searching by Determining Intersections of Search Results (US 6,539,373) - Foundational contextual-search patent: runs multiple sub-queries derived from user context and returns the intersection of their result sets, so context narrows retrieval through set algebra.
- Query Identification and Association (US 8,631,003) - Identifies queries by association patterns across users and sessions, supporting query-canonicalization and intent-cluster identification for the downstream ranking and personalization layers.
Marc Najork, Google Search Patents
- System for Information Extraction from Form-Like Documents (US 11,393,233) - Extracts structured information from form-like documents (invoices, receipts, tax forms) using ML models that combine layout features with text content. Foundational patent for the form-extraction family.
- System for Information Extraction from Form-Like Documents (continuation 2023) (US 11,830,269) - Continuation refining form-extraction logic.
- System for Information Extraction from Form-Like Documents (continuation 2024) (US 12,354,396) - Latest grant in the form-extraction family.
- System for Information Extraction from Form-Like Documents (app 2021) (US App 2021/0374395) - Earliest published application in the form-extraction family.
- System for Information Extraction from Form-Like Documents (app 2022) (US App 2022/0375245) - Continuation application in the form-extraction family.
- System for Information Extraction from Form-Like Documents (app 2024) (US App 2024/0046684) - Continuation application in the form-extraction family.
- System for Information Extraction from Form-Like Documents (app 2025) (US App 2025/0308277) - Latest application in the form-extraction family.
- Systems and Methods for Machine-Learned Prediction of Semantic Similarity Between Documents (US 11,694,034) - Foundational ML-based semantic similarity prediction. Computes per-document embeddings and learned similarity scores that drive modern retrieval, deduplication, and recommendation.
- ML-Learned Semantic Similarity (continuation 2025) (US 12,210,837) - Latest grant in the semantic-similarity family.
- ML-Learned Semantic Similarity (app 2022) (US App 2022/0129638) - Earliest published application in the semantic-similarity family.
- ML-Learned Semantic Similarity (app 2023) (US App 2023/0297783) - Continuation application in the semantic-similarity family.
- ML-Learned Semantic Similarity (app 2025) (US App 2025/0209277) - Latest application in the semantic-similarity family.
- Systems and Methods for Active Learning (US 11,526,752) - Active-learning framework that selects the most informative training examples for human labeling, reducing labeled-data requirements for ML models in retrieval and ranking.
- Systems and Methods for Active Learning (app) (US App 2020/0250527) - Earliest application in the active-learning family.
- Document Activity Logs for Machine Learning (US App 2023/0267277) - Uses document activity logs (clicks, dwell, sustained engagement) to train ML models for document relevance. Cross-listed with the 65 Google Patents collection as pat-61.
- Domain-Based Spam-Resistant Ranking (US App 2007/0067282) - Domain-level ranking signals that resist spam manipulation by reading patterns across the domain rather than per-page, so a single high-quality page on a spam domain cannot pull the domain up.
- Systems and Methods for Ranking Documents Based Upon Structurally Interrelated Information (US App 2005/0060297) - Ranks documents using structural relationships among their constituent information units, capturing inter-relatedness that pure text-match misses.
- Content Evaluation (US App 2006/0069667) - Evaluates content quality using multiple signals including structural features, language quality, and source patterns. A foundational content-quality scoring approach.
- Social Network Recommended Content and Recommending Members for Personalized Search Results (US 8,949,232) - Uses social network signals (who endorsed what, who connected to whom) to personalize search results, surfacing content recommended by trusted social connections.
- Social Network Recommended Content (app) (US App 2013/0086057) - Earliest application in the social-personalization family.
- Using Content Analysis to Detect Spam Web Pages (US App 2006/0184500) - Detects spam web pages through content analysis: reading writing patterns, structural features, and content-quality signals that distinguish spam from legitimate content.
- Systems and Methods for Inferring URL Normalization Rules (US App 2006/0218143) - Automatically infers URL normalization rules so the system canonicalizes URL variants (trailing slashes, parameter orderings, case variations) to one canonical form for indexing and deduplication.
- Automatically Creating Training Data For Language Identifiers (US App 2015/0006148) - Automatically generates training data for language identification models by mining multilingual corpora for known-language samples, scaling language ID training without manual labeling.
- Incremental Update Scheme for Hyperlink Database (US App 2007/0250480) - Incremental update scheme for the hyperlink database that backs web-scale link analysis, supporting fast updates as crawls discover new links without rebuilding the full graph.
- Fault Tolerance Scheme for Distributed Hyperlink Database (US App 2007/0220064) - Fault tolerance scheme for the distributed hyperlink database. Replication, recovery, and consistency mechanisms keep the link graph available at web scale.
Jeffrey Dean, Google Search Patents
- Document Scoring Based on Query Analysis (US 8,051,071) - Foundational ranking patent. Scores documents based on how they relate to query patterns, including query frequency, query refinement patterns, and query-document co-occurrence statistics. Earliest grant in a long continuation family.
- Document Scoring Based on Document Content Update (US 8,112,426) - Scores documents based on patterns of content updates: how often content changes, how substantive the updates are, and how the update pattern correlates with freshness for the query type.
- Document Scoring Based on Document Inception Date (US 7,840,572) - Uses the date a document first appeared on the web as a ranking input, combining inception age with link-acquisition rate. Cross-listed with the 65 Google Patents collection as pat-50.
- Document Scoring Based on Link-Based Criteria (US 8,407,231) - Scores documents using link-based criteria beyond simple link count: link velocity, anchor-text diversity, source-domain quality, and link-burst patterns.
- Document Ranking Based on Document Classification (US 8,224,827) - Ranks documents using their classification into content types and topics, so type-aware ranking can apply different signal weights per document type.
- Information Retrieval Based on Historical Data (US 7,346,839) - Foundational patent introducing historical data as a ranking input: aging, freshness, behavioral history, link history all feed scoring. Cross-listed with the 65 Google Patents collection as pat-51.
- Anchor Tag Indexing in a Web Crawler System (US 7,308,643) - Foundational anchor-text indexing patent. Captures the text used in inbound links to a page and uses it as an additional indexing signal beyond the page content itself.
- Generating Content Snippets Using a Tokenspace Repository (US 8,321,445) - Generates SERP snippets by reading a token-level repository of document content, enabling fast snippet generation that matches the query terms while preserving meaningful context windows.
- Document Compression System and Method for Use with Tokenspace Repository (US 7,917,480) - Document compression scheme tuned for the tokenspace repository: efficient at-rest storage with fast random access by token offset, supporting snippet generation and content retrieval.
- Identifying Related Pages in a Hyperlinked Database (US 6,665,837) - Identifies related pages by analyzing the hyperlink graph, supporting the early "Similar pages" feature and underlying many related-content surfaces.
- Identifying Near-Duplicate Pages in a Hyperlinked Database (US 6,138,113) - Detects near-duplicate pages using fingerprint comparison across the hyperlinked corpus. Foundational deduplication patent for web-scale indexing.
- Finding Mirrored Hosts by Analyzing Connectivity and IP Addresses (US 6,487,555) - Detects mirrored hosts by analyzing IP connectivity patterns and document overlap. Mirrored hosts host the same content; detection prevents duplicate indexing of mirror sites.
- Preventing Topic Drift in Queries in Hyperlinked Environments (US 6,321,220) - Prevents topic drift in query expansion and link-walk-based retrieval. Ensures that expanded queries and link-graph exploration stay on-topic relative to the user query.
- Large Language Models in Machine Translation (US 8,332,207) - Early LLM patent applied to machine translation. Uses large statistical language models trained on huge corpora to score translation candidates, foundational to neural-translation infrastructure.
- Encoding and Adaptive, Scalable Accessing of Distributed Models (US 8,296,123) - Foundational distributed-model serving infrastructure. Encodes large models for efficient distributed storage and adaptive access, supporting model parallelism across many machines.
- System and Method for Efficient Large-Scale Data Processing (US 7,650,331) - The foundational MapReduce patent. Introduces the map-and-reduce paradigm that revolutionized distributed computing and powers Google's crawl, indexing, and ranking infrastructure. Co-invented by Jeffrey Dean and Sanjay Ghemawat.
- System and Method for Analyzing Data Records (US 7,590,620) - Foundational BigTable patent. Distributed structured-data store optimized for huge sparse tables, the substrate of Google's indexing and many search infrastructure systems.
Paul Haahr, Google Search Patents
- Determining Quality of Linked Documents (US 8,825,645) - The flagship Bharat/Singhal/Haahr link-quality patent. Scores documents by affiliation-discounted quality of linking documents - links from independent sources count more than links within tightly affiliated networks.
- Determining Quality of Linked Documents (2012) (US 8,176,056) - Continuation in the link-quality family.
- Determining Quality of Linked Documents (2010) (US 7,783,639) - Earliest grant in the link-quality family.
- Ranking Nodes in a Linked Database Based on Node Independence (US 8,719,276) - Penalizes link networks whose nodes lack independence - the anti-PBN / affiliation-discounting signal. Companion to the link-quality family.
- Determining a Quality Measure for a Resource (US 9,558,233) - The rumored "quality measure" patent. Computes a per-resource quality measure by comparing link popularity to actual user-traffic and engagement signals. Cross-listed with 65gp pat-48; co-invented with Hyung-Jin Kim.
- Document Ranking Based on Document Classification (US 8,224,827) - Classification-as-ranking-signal. Direct ancestor of topical authority / classifier-based reranking. SMX-talk-adjacent.
- Document Scoring Based on Document Inception Date (continuation) (US 8,521,749) - Continuation in the inception-date family. The same patent is 65gp pat-50.
- Information Retrieval Based on Historical Data (US 7,346,839) - Foundational historical-data ranking patent. Aging, freshness, behavioral history, link history. Co-invented with Acharya and Dean. Same patent is 65gp pat-51.
- Document Scoring Based on Query Analysis (2012a) (US 8,185,522) - Continuation in the query-analysis scoring family.
- Document Scoring Based on Query Analysis (2012b) (US 8,239,378) - Continuation in the query-analysis scoring family.
- Document Scoring Based on Query Analysis (2012c) (US 8,244,723) - Continuation in the query-analysis scoring family.
- Document Scoring Based on Query Analysis (2012d) (US 8,266,143) - Continuation in the query-analysis scoring family.
- Document Scoring Based on Query Analysis (2013) (US 8,577,901) - Continuation in the query-analysis scoring family.
- Document Scoring Based on Query Analysis (2014) (US 8,639,690) - Continuation in the query-analysis scoring family.
- Locating Meaningful Stopwords or Stop-Phrases in Keyword-Based Retrieval Systems (US 10,452,718) - Detects which stopwords in a query carry meaning ("the who", "to be") and retains them through retrieval. Foundational query-understanding patent for short-tail intent.
- Locating Meaningful Stopwords (2017) (US 9,817,920) - Continuation in the meaningful-stopwords family.
- Locating Meaningful Stopwords (2015) (US 8,965,919) - Continuation in the meaningful-stopwords family.
- Locating Meaningful Stopwords (2014) (US 8,626,787) - Continuation in the meaningful-stopwords family.
- Locating Meaningful Stopwords (2013) (US 8,473,510) - Continuation in the meaningful-stopwords family.
- Locating Meaningful Stopwords (2012) (US 8,214,385) - Continuation in the meaningful-stopwords family.
- Locating Meaningful Stopwords (2011) (US 7,945,579) - Continuation in the meaningful-stopwords family.
- Locating Meaningful Stopwords (earliest 2008) (US 7,409,383) - Earliest grant in the meaningful-stopwords family.
- Systems and Methods for Providing Search Query Refinements (US 10,223,439) - Generates per-query refinements based on session context, click history, and topical model. Powers the related-searches and refinement chips on SERP.
- Search Query Refinements (2017) (US 9,552,388) - Continuation in the query-refinement family.
- Search Query Refinements (2016) (US 9,495,443) - Continuation in the query-refinement family.
- Search Query Refinements (2014) (US 8,645,407) - Continuation in the query-refinement family.
- Search Query Refinements (2013) (US 8,504,584) - Continuation in the query-refinement family.
- Search Query Refinements (2011b) (US 8,086,619) - Continuation in the query-refinement family.
- Search Query Refinements (2011a) (US 8,065,316) - Earliest grant in the query-refinement family.
- Providing Result-Based Query Suggestions (US 10,459,989) - Generates query suggestions seeded by which results the user clicked, building a click-driven query-rewriting layer.
- Result-Based Query Suggestions (2017) (US 9,563,692) - Continuation in the result-based query suggestions family.
- Result-Based Query Suggestions (2015) (US 9,092,528) - Continuation in the result-based query suggestions family.
- Result-Based Query Suggestions (2013) (US 8,583,675) - Earliest grant in the result-based query suggestions family.
- Multi-Stage Query Processing System and Method for Use with Tokenspace Repository (US 9,146,967) - The canonical multi-stage Google retrieval pipeline: recall → re-rank → final. Operates over the Tokenspace repository.
- Multi-Stage Query Processing (2013) (US 8,407,239) - Earliest grant in the multi-stage query processing family.
- Query Generation Using Structural Similarity Between Documents (US 9,436,747) - Generates expansion queries by comparing structural similarity between documents. Bridges template detection and query rewriting.
- Query Generation Structural Similarity (2015) (US 9,092,479) - Continuation in the query-generation structural-similarity family.
- Query Generation Structural Similarity (2013) (US 8,346,792) - Earliest grant in the query-generation structural-similarity family.
- Document Scoring Content Update (2012) (US 8,234,273) - Continuation in the content-update scoring family.
- Document Scoring Content Update (2013a) (US 8,527,524) - Continuation in the content-update scoring family.
- Document Scoring Content Update (2013b) (US 8,549,014) - Continuation in the content-update scoring family.
- Preferred Sites (US 10,025,868) - Per-user and per-query site preference signals. Captures sitewide trust / brand-preference behaviors and feeds them into ranking.
- Preferred Sites (2016) (US 9,317,563) - Continuation in the preferred-sites family.
- Preferred Sites (2013) (US 8,595,228) - Earliest grant in the preferred-sites family.
- Ranking Based on Reference Contexts (US 8,577,893) - Uses the surrounding context of an anchor (not just the anchor text) as a ranking signal. Modernized anchor-text understanding.
- Propagating Information Among Web Pages (US 8,990,210) - Propagates trust, topicality, and quality signals across site-wide page graphs. Underpins sitewide quality assessment.
- Propagating Information (2013) (US 8,521,717) - Continuation in the propagating-information family.
- Propagating Useful Information Among Related Web Pages (US 7,933,890) - Earliest grant in the propagating-information family.
- Using Synthetic Descriptive Text to Rank Search Results (US 9,208,233) - Generates synthetic descriptive text for images and uses it as a ranking signal. Foundational for image and multimodal ranking.
- Generating Synthetic Descriptive Text (US 9,208,232) - Sibling patent: generates the synthetic descriptive text used by US 9,208,233.
- Generating Descriptive Text for Images Using Seed Descriptors (2018) (US 9,971,790) - Seed-descriptor variant in the synthetic-descriptive-text family.
- Generating Descriptive Text for Images Using Seed Descriptors (2019) (US 10,248,662) - Continuation in the seed-descriptor variant.
- Framework for Evaluating Web Search Scoring Functions (US 8,572,075) - The scorer-evaluation pipeline: A/B-style search-quality testing infra. Directly echoes Haahr's SMX 2016 "how Google works" talk on quality raters and A/B testing of rankers.
- Framework for Evaluating Web Search Scoring Functions (2011) (US 8,060,497) - Earliest grant in the scoring-function-evaluation family.
- Automated Resource Selection Process Evaluation (US 8,489,604) - Quality-rater-feedback-loop infrastructure. Directly relevant to how Google calibrates rankers with the Search Quality Rater Guidelines.
- Embedded Communication of Link Information (US 8,260,766) - Embeds link metadata directly in document structure for ranker consumption. Co-invented with Dean.
- Embedded Communication of Link Information (2011) (US 7,979,417) - Earliest grant in the embedded-link-information family.
- Personally Identifiable Information Detection (US 9,015,802) - Detects PII in indexed content. Co-invented with Matt Cutts.
- Personally Identifiable Information Detection (2013) (US 8,561,185) - Earliest grant in the PII-detection family.
Hyung-Jin Kim, Google Search Patents
- Modifying Search Result Ranking Based on Implicit User Feedback (US 8,661,029) - The foundational Navboost patent. Adjusts search-result rankings based on aggregated implicit user feedback - long clicks, short clicks, skip patterns. Surfaced as the centerpiece of the 2024 DOJ antitrust trial.
- Modifying Search Result Ranking (continuation 2016) (US 9,235,627) - First continuation in the Navboost family.
- Modifying Search Result Ranking (continuation 2017) (US 9,811,566) - Continuation in the Navboost family.
- Modifying Search Result Ranking (continuation 2019) (US 10,229,166) - Continuation in the Navboost family. The one cited most in DOJ-trial coverage.
- Modifying Search Result Ranking (continuation 2021) (US 11,188,544) - Continuation in the Navboost family.
- Modifying Search Result Ranking (continuation 2023) (US 11,816,114) - Most recent continuation in the Navboost family - active IP as of 2023.
- Modifying Search Result Ranking Based on Implicit User Feedback and a Model of Presentation Bias (US 8,938,463) - Sibling to Navboost. Models presentation bias - the tendency for higher-ranked results to receive more clicks regardless of relevance - and corrects for it when deriving relevance signals from clicks.
- Combining User Feedback (US 8,832,083) - Combines primary and secondary user-feedback signals into ranking. Operationalizes the multi-signal click model.
- Accentuating Search Results (US 9,623,119) - Modulates result accentuation based on trust-distribution signals. Complements the Navboost click model.
- Determining a Quality Measure for a Resource (US 9,558,233) - Per-resource quality measure combining link popularity, actual user traffic, and engagement. Co-invented with Paul Haahr. Same patent as 65gp pat-48; canonical article lives in Haahr section.
- Temporal-Based Score Adjustments (US 8,924,379) - Adjusts ranking scores based on temporal patterns - burst detection, seasonality, freshness decay. Complements Dean's content-update family at the per-document score level.
- Modifying Scoring Data Based on Historical Changes (US 8,898,153) - Sibling temporal patent. Adjusts ranking based on historical-change patterns of the document.
- Modifying Ranking Data Based on Document Changes (US 9,002,867) - Adjusts per-document ranking based on the magnitude and pattern of document changes over time. Distinct from temporal-score-adjustments by treating change itself as a per-document signal.
- Grouping of Users (US 8,930,351) - Groups users into cohorts whose collective click signals refine ranking. The cohort-aware variant of Navboost.
- Refining Search Results (US 9,697,259) - Refines results using per-user characteristics. Personalization signal layered on top of Navboost.
- Determining Resource Quality Based on Resource Competition (US 9,020,927) - Scores resource quality by comparing against competing resources for the same query. Per-SERP relative-quality signal.
- Determining Codomain Relationship Measures (US 8,943,099) - Measures relationship between codomains (e.g. document-cluster vs query-cluster). Underpins cross-cluster relevance scoring.
- Ranking Search Results Based on Similar Queries (US 9,009,146) - Borrows ranking signal from semantically similar queries. Foundational for sparse-data ranking - letting popular-query click signals inform rare-query rankings.
- Search Result Inputs Using Variant Generalized Queries (US 9,110,975) - Generalizes queries into variants, then borrows ranking signal across the variants. Companion to similar-query ranking.
- Propagating Query Classifications (US 9,659,097) - Propagates query-classification labels across the query graph. Lets the system classify a long-tail query by analogy to classified head queries.
- Autocompletion Using Previously Submitted Query Data (US 9,740,780) - Autocompletion driven by prior query data. Adjacent to the Matias autocomplete family.
- Enhanced Identification of Interesting Points-of-Interest (US 9,057,616) - POI-quality signal generation. Adjacent to Egnor's local-search prominence work.
- Identifying Landing Pages for Images (US App 2015/0161120) - Identifies the best landing page for an image result. Image-search ranking adjacency.
Pandu Nayak, Google Search Patents
- Integration of Multiple Query Revision Models (US 7,565,345) - Architectural root of the query-revision stack. Integrates multiple per-strategy revision models (synonym, acronym, KHRQ, concept-context) under a unified scoring framework.
- Estimating Confidence for Query Revision Models (US 7,617,205) - The confidence-scoring layer. Per revision candidate, estimates confidence so the system knows when to apply revision and when to leave the query alone.
- Estimating Confidence for Query Revision Models (2012) (US 8,140,524) - Continuation in the confidence-estimation family.
- Estimating Confidence for Query Revision Models (2015) (US 9,069,841) - Latest in the confidence-estimation family. Adds session-based expected-utility scoring.
- Query Revision Using Known Highly-Ranked Queries (US 7,870,147) - KHRQ - revises long-tail queries by borrowing from semantically similar known-high-ranking queries. Click-aggregate-driven query rewriting.
- Query Revision Using Known Highly-Ranked Queries (2013) (US 8,375,049) - Continuation in the KHRQ family.
- Query Revision Using Known Highly-Ranked Queries (app 2011) (US App 2011/0060736) - Continuation application in the KHRQ family.
- Using Concepts as Contexts for Query Term Substitutions (US 9,104,750) - The phrase-aware substitution patent. Treats multi-word phrases as concepts and constrains term substitution by phrasal context - most BERT/RankBrain-adjacent of Nayak's patents.
- Phrase Restricted Substitute Terms (US App 2015/0205866) - Context-locked synonyms. Substitution only applied when the phrase context matches, preventing meaning-shift in ambiguous-term queries.
- Synonym Identification Based on Selected Search Result (US App 2014/0358904) - Pre-RankBrain analog to Navboost. Uses which result a user clicked to learn term substitutes - click signal as synonym-discovery layer.
- Identification of Acronym Expansions (US App 2014/0344263) - Identifies acronym expansions for query understanding. Edge case in query rewriting that materially affects abbreviation-heavy domains.
- Managing and Accessing Data in Web Notebooks (US 8,676,797) - Powered the legacy Google Notebook product. Off the search-ranking spine but documents how Nayak's team thought about user-curated information collections - an idea that later returned in MUM and SGE.
- Managing and Accessing Data in Web Notebooks (app 2007) (US App 2007/0266011) - Earliest application in the Google Notebook family.
Daniel Egnor, Google Local Search Patents
- Scoring Local Search Results Based on Location Prominence (US 8,046,371) - The flagship Local Pack patent. Defines the "Prominence" factor in Google's official Local Search Ranking documentation. Scores local results by location-anchored prominence signals across web mentions, reviews, links, and citations.
- Scoring Local Search Results (app 2011) (US App 2011/0022604) - Pre-grant publication of US 8,046,371.
- Ranking Documents Based on a Location Sensitivity Factor (US 8,171,048) - The "Relevance" factor for location-sensitive queries. Adjusts document ranking by how location-sensitive the query is. Co-invented with Amit Singhal.
- Indexing Documents According to Geographical Relevance (US 9,189,496) - Indexes documents with geographical-relevance metadata so geographic ranking can apply at retrieval time.
- Authoritative Document Identification (US 8,650,197) - Identifies authoritative documents for location-specific queries. The authority signal feeding the Prominence factor.
- Authoritative Document Identification (app 2012) (US App 2012/0173544) - Pre-grant publication of US 8,650,197.
- Classification of Ambiguous Geographic References (US 9,323,738) - Classifies ambiguous geographic references in queries and documents. Handles "Springfield" disambiguation across many U.S. cities sharing the same name.
- Classification of Ambiguous Geographic References (2014) (US 8,856,143) - Earlier grant in the ambiguous-geographic-classification family.
- Classification of Ambiguous Geographic References (app 2015) (US App 2015/0012542) - Continuation application in the ambiguous-geographic-classification family.
- Determining Unambiguous Geographic References (US 8,078,601) - Companion to the ambiguous-geographic-classification family. Resolves geographic references that have a single unambiguous mapping.
- Processing Ambiguous Search Requests in a Geographic Information System (US 10,198,521) - Processes ambiguous search requests in Google Maps. Co-invented with John Hanke, founder of Keyhole and CEO of Niantic.
- Processing Ambiguous Search Requests in GIS (app 2019) (US App 2019/0171688) - Continuation application in the GIS ambiguous-search family.
- Processing Ambiguous Search Requests in GIS (app 2015) (US App 2015/0169674) - Earlier application in the GIS ambiguous-search family.
- Category Suggestions Relating to a Search (US 8,595,250) - Suggests business categories for local listings and search refinement. The category-resolution layer that powers GBP category-aware ranking.
- Category Suggestions Relating to a Search (2011) (US 8,027,988) - Earliest grant in the category-suggestions family.
- Propagating Information Among Web Pages (US 8,990,210) - Propagates trust, topicality, and quality signals across site-wide page graphs. Egnor is co-inventor; canonical article lives in Haahr's section.
- Propagating Information Among Web Pages (2013) (US 8,521,717) - Continuation in the propagating-information family.
- Propagating Useful Information Among Related Web Pages (US 7,933,890) - Earliest grant in the propagating-information family.
- Propagating Information Among Web Pages (app 2011) (US App 2011/0196861) - Continuation application in the propagating-info family.
- Propagating Information Among Web Pages (app 2014) (US App 2014/0052735) - Continuation application in the propagating-info family.
Yossi Matias, Google SERP Features & Assistant Patents
- Generating Related Questions for Search Queries (US 9,213,748) - The foundational People-Also-Ask patent. Generates per-query related-question candidates from query logs, entity models, and topical co-occurrence - the system that powers the PAA SERP feature.
- Generating Related Questions for Search Queries (2017) (US 9,679,027) - Continuation in the People-Also-Ask family.
- Generating Additional Content (US App 2016/0026727) - Generates supplemental SERP content. Adjacent to PAA; produces inline content blocks that expand the SERP beyond the ten blue links.
- Query Formulation and Search in the Context of a Displayed Document (US 9,342,601) - Formulates and runs search queries in the context of a displayed document. Predictive-search and contextual-query DNA.
- Third Party Search Applications for a Search System (US 10,289,618) - Surfaces third-party search applications inline on the SERP. The structural primitive for federated search and OneBox integration.
- Identifying Central Entities (US 9,009,192) - Identifies the central entities of a query, document, or corpus. The scaffolding for Knowledge Panel and entity-aware ranking.
- Grouping Related Entities (US 9,411,857) - Groups related entities into clusters. Companion to central-entity selection.
- Providing an Explanation of a Missing Fact Estimate (US 10,318,540) - When a Knowledge Panel fact is missing or estimated, generates an explanation. The transparency layer for fact-panel reasoning.
- Selecting Textual Representations for Entity Attribute Values (US 10,685,073) - Selects the best textual representation for entity attributes. Drives how Knowledge Panel and entity cards render attribute values.
- Searching Quotes of Entities Using a Database (US 10,198,508) - Searches and surfaces quotes attributed to entities. Powers the entity-quote SERP feature.
- Campaign and Competitive Analysis and Data Visualization Based on Search and Social Media (US 9,043,302) - Analyzes campaigns and competitive landscape from search and social-media data. The analytics-adjacent precursor to Trends-driven insights.
- Ranking Graphical Visualizations of a Data Set According to Data Attributes (US 8,812,947) - Ranks graphical visualizations of a dataset by attribute fit. The visualization-ranking primitive for Trends-style dashboards.
- Advertising Inventory Allocation (US 8,615,436) - Allocates advertising inventory across surfaces. Ad-systems infrastructure.
- Automated Call Requests with Status Updates (US 10,827,064) - The Google Duplex automated-calling patent. Places autonomous phone calls on the user's behalf with conversational voice and status updates.
- Automated Call Requests with Status Updates (app 2019) (US App 2019/0306314) - Earlier application in the Duplex automated-calls family.
- Automated Call Requests (app 2024) (US App 2024/0187521) - Continuation application in the Duplex automated-calls family.
- Reducing Telephone Network Traffic Through Pre-Call Information (US App 2024/0205331) - Pre-call information mechanism for Duplex.
- Synchronous Communication Using Voice and Text (US 11,811,968) - Synchronous voice + text communication. The multimodal communication primitive behind Assistant.
- Synchronous Communication Using Voice and Text (2025) (US 12,301,754) - Continuation in the synchronous voice/text family.
- System(s) and Method(s) for Implementing a Personalized Chatbot (US 12,225,158) - Implements per-user personalized chatbot. The modern Assistant-LLM era patent.
- Personalized Chatbot (app 2025) (US App 2025/0150532) - Continuation application in the personalized chatbot family.
- Large Language Model Output Entailment (US App 2025/0094456) - Validates LLM-generated output via entailment checking against grounding sources. The grounding/RAG-validation layer for Gemini / Assistant LLM outputs.
- Generating a Composite Response to Natural Language Input (US App 2025/0069597) - Generates composite multi-modal responses to natural-language input. Modern Assistant generative response composition.
- Selectively Generating Expanded Responses That Guide Continuance of a Human-to-Computer Dialog (US App 2022/0157308) - Selectively expands responses to guide continuance of dialog. The proactive Assistant turn-guidance primitive.
- Audio-Visual Hearing Aid (US App 2024/0428816) - Audio-visual hearing-aid system. Accessibility primitive integrating speech recognition and visual context.
- Audio-Visual Hearing Aid (app 2023) (US App 2023/0267942) - Earlier application in the hearing-aid family.
- Using Video Clips as Dictionary Usage Examples (US App 2025/0173509) - Uses video clips as dictionary usage examples. The Dictionary OneBox modernization for multimodal context.
- Text-to-Speech Interface Featuring Visual Content Supplemental to Audio Playback (WO2020023070A1) - TTS interface with supplemental visual content. Multimodal Assistant interface primitive.
- Aggregation of Related Media Content (US 11,335,380) - Aggregates related media content. Assistant content-collection primitive.
- Aggregation of Related Media Content (2020) (US 10,770,112) - Earlier grant in the aggregation-media family.
- Automatic Execution of Actions by a Mobile Computing Device (DE112019000305T5) - Automatic action execution by mobile device. Assistant on-device action primitive.
- Dynamically Updating User Interface Elements (EP3827332A1) - Dynamically updates UI elements. Assistant interface primitive.
- Neural Networks with Selective Attention Layers (WO2026044251A1) - Selective attention layers in neural networks. ML-architecture primitive.
- Text-Driven Image Editing via Image-Specific Finetuning of Diffusion Models (EP4581581A1) - Image-specific finetuning of diffusion models for text-driven editing.
- Text-Driven Image Editing (WO) (WO2024086598A1) - WO publication of the text-driven image editing patent.
- Conditioning Text-to-Image Generative Neural Networks (WO2024249439A1) - Conditioning text-to-image generative neural networks.
- Video Editing Using Diffusion Models (WO2024163624A1) - Video editing via diffusion models.
- Methods and Systems for Fast Inference from Machine Learning Models (WO2024118603A1) - Fast inference techniques for ML models. Serving-infrastructure primitive.
- Partially Customized Machine Learning Models for Data De-identification (US App 2021/0303725) - Partially customized ML models for data de-identification. Privacy-engineering primitive.
- Generating Contact Suggestions (US 9,934,286) - Generates contact suggestions for social and communication apps.
- Generating Contact Suggestions (2016) (US 9,311,415) - Earlier grant in the contact-suggestions family.
- Systems and Methods for Providing Contact Group Member Suggestions (US 9,773,229) - Contact group member suggestions.
- Generating Suggestions for User Groups in Social Networks (US 8,977,649) - Suggests user groups in social networks.
- Generating Suggestions for User Groups (2014) (US 8,880,607) - Earlier grant in the social user-groups family.
- Security for Wireless Broadcasts (US 10,616,765) - Security primitive for wireless broadcasts. Eddystone-adjacent.
- Security for Wireless Broadcasts (EP) (EP3320711A1) - EP publication of the wireless-broadcasts security patent.
- Panoramic Video Acquisition Guidance (US App 2015/0124047) - Panoramic video acquisition guidance. Photography/capture primitive.
Andrei Broder, Cross-Vendor Search Patents
- Method for Estimating the Probability of Collisions of Fingerprints (US 5,974,481) - Detects near-collisions among fingerprints of data strings via one-way function plus bit-masking. Estimates exact-collision probability without storing multiple fingerprints. The DEC-era foundational patent for fingerprinting infrastructure.
- Method for Clustering Closely Resembling Data Objects (MinHash / Shingling) (US 6,349,296) - The foundational MinHash patent. Partitions documents into shingles, fingerprints them, applies pseudo-random permutations to generate sketches, clusters objects sharing sufficient features. Efficient near-duplicate detection at web scale. Cited across every modern dedup system.
- Method and Apparatus for Ranking Web Page Search Results (US 6,560,600) - The AltaVista ranking patent. Ranks linked-database search results using linear combinations of matrices and eigenvector analysis. Incorporates attractor/non-attractor matrices, probability weighting, co-citation, bibliographic coupling.
- Method for Clustering Closely Resembling Data Objects (DEC-era) (US 6,119,124) - DEC-era MinHash/shingling sibling to US 6,349,296. Earlier patent in the family.
- Method for Determining the Resemblance of Documents (US 6,230,155) - Canonical shingling / Jaccard sketch patent. Document-resemblance estimation via shingled fingerprints - the formal complement to MinHash clustering.
- Compression Protocol with Multiple Preset Dictionaries (US 5,953,503) - Network compression with shared dictionaries. DEC-era foundational infrastructure (Broder + Mitzenmacher + Mogul).
- Method and Apparatus for Ranking Web Page Search Results (Overture 2005) (US 6,871,202) - Overture-era eigenvector ranking continuation in the AltaVista-ranking family.
- Method for Ranking Web Page Search Results (Overture 2008) (US 7,398,461) - Overture SERP-ranking continuation (Broder + Maghoul).
- UIMA / Unstructured Information Management with Multi-Tokenization Views (US 7,139,752) - IBM UIMA backbone - multi-tokenization views for unstructured information management. Foundational for IBM enterprise search.
- Method for Identifying Related Pages in a Hyperlinked Database (Yahoo) (US 7,630,973) - Related-pages / co-citation continuation co-invented with Monika Henzinger. Cross-listed in Henzinger's section.
- Methods and Apparatus for Assessing Web Page Decay (US App 2008/0097978) - Assesses web page staleness by examining internal date references and other temporal indicators within page content. The IBM-era patent that adds a per-page decay signal complementing the Dean/Haahr freshness families.
- Methods and Apparatus for Assessing Web Page Decay (sibling app) (US App 2008/0097977) - Sibling application to US App 2008/0097978 in the web-page decay family.
- Ad Matching by Augmenting a Search Query with Knowledge Obtained Through Search (US App 2009/0254512) - Matches advertisements to search queries by producing ad queries that include unigram features and external classification features. Bridges search-quality signals into ad relevance.
- Context Transfer in Search Advertising (US 8,886,636) - Classifies landing pages using trained classifiers based on search queries. Determines appropriate page types for web searchers - the search-to-landing-page intent bridge.
- System and Method for Providing Contextual Actions on a Search Results Page (US 9,015,140) - Targeted applications within SERP. Interprets queries and maps to relevant inline contextual actions. The structural primitive for inline SERP-feature integration.
- Method and System for Quantifying User Interactions with Web Advertisements (US 8,812,362) - Probability prediction for web advertisement clicks. Expert statistical models trained on web document features determine click probability. The click-modeling cousin of Kim's Navboost click signal applied to ads.
- Method and System for Using Email Receipts for Targeted Advertising (US App 2012/0047014) - User classification based on email content analysis for targeted advertising.
- Presentation of Content Based on Utility (US App 2012/0084155) - Presents articles based on utility metrics with preview-icon links. Content-selection method for utility-driven surfaces.
- Method for Selectively Restricting Access to Computer Systems (Foundational CAPTCHA) (US 6,195,698) - The foundational CAPTCHA patent. Selectively restricts access to computer systems by challenge-response tests that humans can solve but bots cannot. Co-invented with Mark Lillibridge, Martin Abadi, and Krishna Bharat - one of the landmark security/anti-bot patents in computing.
- Multi-Step CAPTCHA with Serial Time-Consuming Decryption of Puzzles (US 8,522,327) - Challenge-response security testing with sequential puzzles. Users must solve each challenge before subsequent challenges are revealed. Continues Broder's long-running CAPTCHA-design lineage.
- Generating Hard Instances of CAPTCHAs (US App 2010/0077209) - Yahoo-era CAPTCHA continuation. Generates hard instances to defeat bot solvers.
- CAPTCHA Image Generation (US App 2010/0077210) - Yahoo-era CAPTCHA image-generation continuation.
- Classification and Storage of Events in a Network (US 8,666,819) - Yahoo computational-advertising event-store backbone. Classifies and stores network events at scale.
- System and Method for a Cloud-Based Electronic Communication Vault (US 8,788,819) - Secure storage and delivery of information via cloud-based vault. Secure data repository for authorized users.
- System and Method for Social Filtering of Comments (US 12,039,493) - Sorts comments by commenter social status. Recent (2024) Yahoo Assets work on social-content ranking surfaces.
Monika Henzinger, Google Search Patents
- Detecting Duplicate and Near-Duplicate Files (US 6,658,423) - Henzinger-led near-duplicate file detection with William Pugh. Operates on arbitrary files (not just web pages) using hashing and clustering. The structural pattern for canonicalization at file level.
- Detecting Duplicate and Near-Duplicate Files (2008) (US 7,366,718) - Continuation in the duplicate-files family.
- Detecting Duplicate and Near-Duplicate Files (2011) (US 8,015,162) - Continuation in the duplicate-files family.
- Detecting Duplicate and Near-Duplicate Files (2016) (US 9,275,143) - Most recent grant in the duplicate-files family.
- System and Method for Near-Uniform Sampling of Web Page Addresses (US 6,594,694) - Near-uniform sampling of web-page addresses for crawl quality assessment.
- Connectivity Server for Locating Linkage Information Between Web Pages (US 6,073,135) - The AltaVista Connectivity Server - foundational link-graph infrastructure. Stores and serves linkage information between web pages efficiently. The structural ancestor of every modern link-graph backend.
- Method for Ranking Hyperlinked Pages Using Content and Connectivity Analysis (US 6,738,678) - Combines content analysis with link-graph connectivity for hyperlinked-page ranking. Pre-PageRank-era multi-signal ranking primitive that influenced later combined-signal ranking models.
- Method for Ranking Documents in a Hyperlinked Environment Using Connectivity and Selective Content Analysis (US 6,112,203) - Earliest grant in the connectivity-and-content-ranking family.
- Method for Ranking Hyperlinked Pages (continuation 2007) (US 7,117,206) - Continuation in the connectivity-and-content-ranking family.
- Method for Identifying Related Pages (2009) (US 7,630,973) - Continuation in the related-pages family.
- Method and Apparatus for Finding Mirrored Hosts by Analyzing URLs (US 6,286,006) - URL-based variant of mirrored-host detection.
- Ranking Search Engine Results (US 7,451,388) - Ranking search engine results based on user-behavior signals.
- Methods and Systems for Identifying Manipulated Articles (US 7,302,645) - Detects manipulated articles in web corpora. Spam-detection adjacent.
- Identification of Web Sites That Contain Session Identifiers (US 7,886,217) - Identifies web sites that contain session identifiers in URLs. URL-canonicalization infrastructure.
- Document Ranking Based on Document Classification (US 8,224,827) - Document-classification ranking. Henzinger is co-inventor. Canonical article in Haahr section.
- Systems and Methods for Determining Document Freshness (US 7,797,316) - Determines per-document freshness from temporal signals. Distinct from Dean's content-update family by focusing on freshness assessment rather than update velocity.
- Systems and Methods for Determining Document Freshness (2011) (US 8,082,244) - Continuation in the document-freshness family.
- Systems and Methods for Determining Document Freshness (2013) (US 8,515,952) - Continuation in the document-freshness family.
- Document Ranking Based on Semantic Distance Between Terms in a Document (US 7,716,216) - Ranks documents by semantic distance between query terms within the document. The pre-neural-era semantic-proximity signal.
- Document Ranking Based on Semantic Distance (2011) (US 8,060,501) - Continuation in the semantic-distance ranking family.
- Document Ranking Based on Semantic Distance (2013) (US 8,606,778) - Continuation in the semantic-distance ranking family.
- Search Queries Improved Based on Query Semantic Information (US 8,055,669) - Improves search queries using semantic information about the query itself. Pre-RankBrain query-understanding primitive.
- Search Queries Improved Based on Query Semantic Information (2013) (US 8,577,907) - Continuation in the query-semantic-information family.
- Systems and Methods for Using Anchor Text as Parallel Corpora for Cross-Language Information Retrieval (US 7,146,358) - Uses anchor text as parallel corpora for cross-language information retrieval. Foundational for multilingual search and translation-quality ranking.
- Anchor Text as Parallel Corpora (2010) (US 7,814,103) - Continuation in the anchor-text cross-language IR family.
- Anchor Text as Parallel Corpora (2014) (US 8,631,010) - Continuation in the anchor-text cross-language IR family.
- Systems and Methods for Performing In-Context Searching (US 7,305,380) - Performs in-context searching - searches initiated from current user context (document, page, location). The structural ancestor of Search Lens and contextual lookup.
- In-Context Searching (2011) (US 7,962,469) - Continuation in the in-context searching family.
- In-Context Searching (2015) (US 9,111,000) - Continuation in the in-context searching family.
- In-Context Searching (2017) (US 9,665,650) - Latest grant in the in-context searching family.
- Hypertext Browser Assistant (US 7,421,432) - Browser-assistant infrastructure for in-browser search and content lookup. Pre-Chrome integration patent foreshadowing modern browser-search integration.
- Hypertext Browser Assistant (2012) (US 8,316,016) - Continuation in the browser-assistant family.
- Hypertext Browser Assistant (2013) (US 8,560,564) - Continuation in the browser-assistant family.
- Methods and Apparatus for Employing Usage Statistics in Document Retrieval (US 8,001,118) - Uses aggregate usage statistics to inform document retrieval. The retrieval-layer ancestor of click-driven ranking signals that later became Navboost-style aggregations.
- Usage Statistics Document Retrieval (2012) (US 8,156,100) - Continuation in the usage-statistics retrieval family.
- Usage Statistics Document Retrieval (2013) (US 8,352,452) - Continuation in the usage-statistics retrieval family.
- Voice Interface for a Search Engine (US 7,027,987) - Voice-driven interface for search-engine queries. Foundational voice-search infrastructure.
- Voice Search Engine Interface for Scoring Search Hypotheses (US 8,768,700) - Scores search hypotheses for voice-driven search. Voice-query disambiguation.
- Ranking Video Articles (US 7,933,338) - Ranks video articles for video-search results.
- Finding Web Pages Relevant to Multimedia Streams (US 8,868,543) - Finds web pages relevant to multimedia streams.
- Methods and Systems for Improving Search Rankings Using Advertising Data (US 8,676,790) - Improves search rankings using advertising data signals. Ad-to-organic cross-signal infrastructure.
Trystan Upstill, Google Site-Quality & HCU-Era Search Patents
- Selectively Generating Alternative Queries (Panda) (US 9,135,307) - The Panda patent. Generates alternative queries to suppress low-quality sites. Cross-listed with 65gp pat-39 (Panda). Co-invented with Navneet Panda and April Lehman.
- Re-ranking Resources Based on Categorical Quality (US App 2019/0155948) - The HCU-era follow-up to Panda. Re-ranks results by categorical site-quality assessment. Lead-inventor patent.
- Obtaining Authoritative Search Results (US 9,659,064) - Identifies authoritative search results via source-quality signals. The authority-focused complement to categorical quality.
- Ranking Search Results (US 9,454,582) - Per query, ranks search results by combined site- and content-quality signals.
- Selectively Ranking Search Results (US 9,442,990) - Per query and per candidate result, selectively applies ranking based on result-eligibility criteria.
- Resource Identification from Organic and Structured Content (US 9,589,028) - Identifies canonical resources by combining organic-content signals with structured-data markup. Bridges Schema.org markup into resource identification.
- Search Operation Adjustment and Re-Scoring (US 10,339,144) - Adjusts search operations and re-scores results based on quality signals discovered during retrieval.
- Identifying Affiliated Domains (US 9,178,848) - Detects affiliated domains across ownership, hosting, content, and link patterns. The structural anti-PBN domain-cluster detection signal.
- Determining Resource Attributes from Site Address Attributes (US 8,600,993) - Extracts resource attributes (topic, type, locale) from site-address structural attributes. URL-pattern-derived signals.
- Navigational Resources for Queries (US 9,213,774) - Identifies navigational-intent resources for queries. The navigational-query handling primitive.
- Navigational Resources for Queries (continuation 2019) (US 10,204,138) - Continuation in the navigational-resources family.
- Identifying Document Languages Using Link Context (US 9,098,582) - Identifies document language via the link context surrounding inbound anchors. Multilingual retrieval primitive.
- Identifying Document Languages (continuation 2019) (US 10,223,461) - Continuation in the document-language family.
- Identifying Points of Interest (US 8,239,130) - POI-identification primitive for local search.
- Identifying Points of Interest (2013) (US 8,433,512) - Continuation in the POI-identification family.
- Identifying Points of Interest (2013b) (US 8,620,579) - Continuation in the POI-identification family.
- Identifying Implicitly Local Queries (US 8,200,694) - Identifies queries with implicit local intent without explicit location terms.
- Encoding Distances Between Locations (US 8,495,046) - Efficient distance encoding for local search infrastructure.
- Geographic Synonyms (US 8,417,721) - Synonyms for geographic references (e.g., NYC = New York City) for local-query handling.
- Restricted-Locality Synonyms (US 8,719,282) - Synonyms restricted by locality. Per-locale synonym applicability for query rewriting.
- Lexical Synonyms (US 9,183,297) - Lexical-synonym identification for query rewriting.
- Query Correction (US 9,378,272) - Query-correction infrastructure (spell, intent).
- Related Terms Across Languages (US 8,798,988) - Identifies related terms across languages for cross-language IR.
- Phrase Restricted Substitute Terms (US App 2015/0205866) - Phrase-restricted substitute terms. Co-invented with Pandu Nayak - cross-list to Nayak section.
- Locally Significant Queries (US App 2014/0172843) - Identifies queries locally significant in a region but not globally significant.
- Co-Occurrence Based List Extraction (US 8,285,738) - Extracts lists from co-occurrence patterns in web content.
- Co-Occurrence List Extraction (2016) (US 9,239,823) - Continuation in the co-occurrence list-extraction family.
- Using Language Models to Generate User Interface Components (WO 2026/072958) - First LLM-era Upstill filing. Uses LLMs to generate UI components dynamically based on query/context.
Susan Dumais, Microsoft IR & Search Patents
- Computer Information Retrieval Using Latent Semantic Structure (LSI) (US 4,839,853) - The foundational Latent Semantic Indexing patent. Uses singular value decomposition to capture latent semantic relationships between documents and queries - the conceptual ancestor of dense-embedding retrieval systems. Co-invented with Deerwester, Furnas, Harshman, Landauer, Lochbaum, Streeter.
- Automated Satisfaction Measurement for Web Search (US 7,937,340) - Automated measurement of SERP satisfaction from user signals. The Microsoft parallel to Google's Navboost click-driven ranking infrastructure.
- Preference Judgements for Relevance (US 8,069,179) - Pairwise preference-judgment click model for relevance signal. Modern learning-to-rank foundation.
- Cursor Activity Evaluation for Search Result Enhancement (US App 2013/0246383) - Uses cursor activity (hover, dwell, movement) beyond clicks as implicit-feedback signal. Modern multi-modal interaction signal.
- Relating Web Page Change with Revisitation Patterns (US 8,078,974) - Relates web-page change patterns to user revisitation. The freshness × behavior cross-signal - pages users return to as they change carry distinct relevance signal.
- Web Page Change × Revisitation (2015) (US 9,069,872) - Continuation in the page-change × revisitation family.
- Modeling Intent and Ranking Search Results Using Activity-Based Context (US App 2012/0158685) - Models user intent and ranks results using activity-based context. Modern context-aware ranking primitive.
- Personalized Navigation Using a Search Engine (US 8,799,280) - Per-user personalized navigation through search engine. Personalization-driven result re-ranking.
- Functionality for Personalizing Search Results (US 8,700,544) - Functionality and architecture for per-user search-result personalization.
- Domain Expertise Determination (US 8,122,021) - Per user, models domain expertise across topics. The user-side authority/expertise model - complement to Google's document-side authority work.
- Domain Expertise (2013) (US 8,402,024) - Continuation in the domain-expertise family.
- Domain Expertise (2015) (US 8,930,357) - Continuation in the domain-expertise family.
- Authority Ranking (US 8,260,789) - Authority-ranking weight and confidence functions. Microsoft's formal authority-ranking framework.
- Time-Aware Ranking Adapted to a Search Engine Application (US 9,244,931) - Time-aware ranking with adaptive freshness sensitivity. Microsoft's temporal ranking primitive complementing Google's Dean/Haahr freshness families.
- Time-Aware Ranking (2019) (US 10,346,413) - Continuation in the time-aware ranking family.
- Assigning Relevance Weights Based on Temporal Dynamics (US 10,353,967) - Assigns relevance weights based on temporal dynamics - how relevance shifts over time per topic.
- Automatic Identification and Contextual Reformulation of Implicit Device-Related Queries (US 11,386,105) - Automatic detection and contextual reformulation of implicit device-related queries. Modern multi-device query understanding (Adam Fourney co-inv).
- Using Categorical Metadata to Rank Search Results (US 9,020,936) - Uses categorical metadata to rank results. Category-aware ranking primitive.
Christopher Burges, Microsoft Research Learning-to-Rank Patents
- Ranking Results Using Multiple Nested Ranking (RankNet) (US 7,689,615) - The RankNet patent. Pairwise neural ranking learned via gradient descent on a probabilistic cost function. Foundational learning-to-rank patent that started the lineage to LambdaRank and LambdaMART.
- Efficiency of Training for Ranking Systems Based on Pairwise Training with Aggregated Gradients (LambdaRank) (US 7,617,164) - The LambdaRank patent. Pairwise gradient aggregation enables training with arbitrary cost functions including NDCG-aligned objectives. The structural innovation that lets ranking systems train directly on what evaluation metrics actually measure.
- Training a Learning System with Arbitrary Cost Functions (LambdaRank app) (US App 11/305,395) - Earliest LambdaRank application - arbitrary cost function training including direct NDCG optimization.
- Boosting a Ranker for Improved Ranking Accuracy (LambdaMART) (US App 12/032,697) - The LambdaMART patent. Gradient-boosted ensemble of LambdaRank trees - the model that won the Yahoo Learning to Rank Challenge in 2010 and underpins gradient-boosted ranking at Bing and many other engines.
- Linear Combination of Rankers (US App 12/836,660) - Ensemble integration of multiple rankers via linear combination. Production ranking systems use ensembles of LambdaMART models; this patent covers the combination layer.
- Linear Combination of Rankers (app 2007) (US App 11/975,518) - Earlier application in the ranker-combination family.
- System and Method for Learning Ranking Functions on Data (US App 11/462,932) - Broader ranking-function family - generalizes RankNet/LambdaRank approach.
- Accounting for Behavioral Variability in Web Search (US App 11/904,103) - Models behavioral variability in click logs. Adapts ranking to handle the noise in implicit-feedback signals - the Microsoft parallel to Kim's presentation-bias modeling.
- Smooth Link-Spam Classification (US App 13/921,862) - Smooth (continuous-probability) link-spam classification. The structural alternative to hard binary spam-classification approaches; uses smoothing across link-graph neighborhoods.
- Link-Spam Smoothing (2008 app) (US App 11/901,072) - Earlier application in the link-spam smoothing family.
- Web-Page Analysis Multi-Graph (US App 11/901,049) - Multi-graph analysis of web pages combining link, content, and behavioral graphs for spam-resistant ranking.
- Re-Ranking Top Results (US App 12/421,022) - Re-ranks the top results using additional ranker stage. Mid-pipeline re-ranking primitive.
- Margin-Based Tree Classification (US App 10/179,049) - Margin-tree classification - boosted tree ensemble with margin-based training objective. Pre-LambdaMART boosting infrastructure.
- K-Nearest-Neighbor Probability Estimation (US App 11/296,919) - k-NN probability estimation for classification and ranking.
- Semi-Supervised Graphical Model Training (US App 11/170,989) - Semi-supervised training for graphical models. Foundation for label-efficient ranking model training.
- Auto Playlist Generator (US App 09/870,292) - Earliest Burges patent. Audio-feature-based playlist generation. Peripheral to search but foundational for his ML infrastructure work.
Eric Brill, Microsoft Research Search & NLP Patents
- Systems and Methods for Improved Spell Checking (US 7,254,774) - The foundational noisy-channel query speller patent. Models spell correction as inferring the intended query from observed noisy input via probabilistic channel model. Underpins "did you mean" surfaces across every modern search engine.
- Systems and Methods for Improved Spell Checking (app) (US App 2007/0106937) - Continuation application in the noisy-channel spell-check family.
- Spell Checker with Arbitrary Length String-to-String Transformations (US 7,366,983) - Spell checker using arbitrary-length string-to-string transformations to improve noisy-channel spelling correction. Captures multi-character substitutions, insertions, deletions.
- Spell Checker String-to-String (2007) (US 7,290,209) - Earlier grant in the string-to-string spell-checker family.
- Automated Error Checking System and Method (US App 2007/0016616) - General automated error-checking framework. Generalizes the spell-check approach to broader error categories.
- Accounting for Behavioral Variability in Web Search (US 7,743,047) - Behavioral-variability accounting in web search ranking. Click signals vary by user - this patent models that variability for more robust click-driven ranking.
- Accounting for Behavioral Variability (app) (US App 2008/0281817) - Continuation application in the behavioral-variability family.
- Using Popularity Data for Ranking (US App 2007/0100824) - Uses popularity data signals (visits, page rank, social signals) to inform document ranking.
- Mining Web Search User Behavior to Enhance Web Search Relevance (US App 2007/0208730) - Mines aggregate user behavior across web search to enhance relevance. Pre-Navboost-era click-driven ranking primitive.
- Page-Biased Search (US App 2006/0242138) - Per-page bias signals applied to search ranking. Specific pages can be biased up or down based on quality signals.
- Using Connectivity Distance for Relevance Feedback in Search (US App 2007/0239702) - Uses connectivity distance in link graph as relevance feedback signal.
- Cost-Benefit Approach to Automatically Composing Answers to Questions (US App 2006/0294037) - Cost-benefit framework for automatically composing answers to questions by extracting information from large unstructured corpora. Pre-LLM-era Q&A primitive.
- Question Answering Over Structured Content on the Web (US App 2007/0094285) - Q&A using structured web content as knowledge source.
- User Intent Discovery (US App 2007/0162442) - Discovers user intent from query patterns and behavior.
- Reducing Human Overhead in Text Categorization (US App 2007/0183655) - Reduces human labeling overhead in text categorization via active learning and bootstrapping.
- Arbitration of Specialized Content Using Search Results (US App 2007/0078822) - Arbitrates which specialized vertical search results to surface on a general SERP.
- Location-Aware Multi-Modal Multi-Lingual Device (US App 2007/0005363) - Location-aware multi-modal multi-lingual device infrastructure.
- Associating Information with an Electronic Document (US App 2006/0242574) - Associates external information with electronic documents.
- Associating Supplementary Information with Network-Based Content Locations (US App 2006/0224662) - Associates supplementary information with network-based content locations (URLs).
- Intelligent Integration of Notifications in Mobile Health Systems (US App 2017/0308650) - Late application - mobile-health notifications. Peripheral to search.
Tomas Mikolov, Google word2vec Patents
- Computing Numeric Representations of Words in a High-Dimensional Space (word2vec) (US 9,037,464) - The foundational word2vec patent. Learns continuous numeric representations of words in a high-dimensional vector space such that semantically and syntactically related words are nearby. The 2013 architecture (CBOW and Skip-gram) is the conceptual root of every dense-embedding NLP model since.
- Computing Numeric Representations of Words (continuation) (US 9,740,680) - Continuation of the word2vec patent broadening claim scope. Same disclosure as US 9,037,464.
Noam Shazeer, Google Transformer, MoE & Search Patents
- Attention-Based Sequence Transduction Neural Networks (Transformer) (US 10,452,978) - The foundational Transformer patent. Introduces self-attention as the primary mechanism for sequence transduction, replacing recurrence and convolution. The 2017 architecture underpinning BERT, GPT, T5, PaLM, Gemini, and every modern large language model.
- Attention-Based Sequence Transduction (2020 continuation) (US 10,719,764) - First continuation in the Transformer family.
- Attention-Based Sequence Transduction (2021) (US 11,113,602) - Continuation in the Transformer family.
- Attention-Based Sequence Transduction (2025) (US 12,217,173) - Latest grant in the Transformer family.
- Decoder-Only Transformer Architecture (US 12,354,005) - Decoder-only Transformer variant. Underpins GPT-style autoregressive language models.
- Attention-Based Image Generation (US 12,142,034) - Attention-based image generation. Multimodal application of attention mechanism. Underpins image-search and Lens-style multimodal retrieval.
- Speech-Recognition Attention RNN (US 12,100,391) - Speech-recognition attention RNN. Attention mechanism applied to speech.
- Mixture of Experts Neural Networks (US 11,769,055) - The sparsely-gated Mixture-of-Experts patent. Routes inputs through experts via learned gating, enabling neural networks to scale parameters without proportional compute increase. Co-invented with Jeff Dean, Geoffrey Hinton, Quoc Le, Azalia Mirhoseini.
- Mixture of Experts (2024) (US 12,067,476) - Continuation in the MoE family.
- Neural Networks with Switch Layers (Switch Transformer) (US 12,093,829) - The Switch Transformer patent. Simplified MoE routing using a single-expert-per-token gating mechanism. Scales to trillion-parameter models with manageable compute.
- Using Large Language Models in Generating Automated Assistant Responses (US 12,148,421) - Uses LLMs to generate automated assistant responses. The SGE / AI-Overviews lineage patent - generates natural-language search responses grounded in retrieved content.
- Evaluating Output Sequences Using an Auto-Regressive Language Model (US 12,086,713) - Evaluates output sequences using an auto-regressive LM. The LLM-as-judge / quality-scoring pattern - uses LLMs to evaluate other LLMs' outputs.
- Distributing Tensor Computations (Pathways / GSPMD) (US 12,265,903) - Distributes tensor computations across accelerators. The Pathways/GSPMD infrastructure that makes large-model training feasible.
- Modifying Search Result Ranking Based on Implicit User Feedback (Navboost) (US 8,661,029) - The Navboost patent. Shazeer is co-inventor with Kim, Tong, Diligenti. Canonical article in Kim section.
- Ranking Documents Based on Large Data Sets (US 9,116,976) - Large-scale document ranking model. Pre-Transformer-era large-data-set ranking infrastructure (with Bem, Harik, Tong).
- Model Generation for Ranking Documents (US 7,743,050) - Model-generation infrastructure for document ranking. Earlier in the large-data-set ranking lineage.
- Equivalent Descriptions for an Information Need (US 7,392,244) - Synonyms / equivalent-descriptions query rewriting precursor. Early Shazeer ranking work.
- Determining Geographical Relevance of Web Documents (US 8,086,690) - Determines geographical relevance of web documents. Adjacent to Egnor's local-search work.
Ashish Vaswani, Google Transformer, Attention-Based Vision & Fast Decoding Patents
- Attention-Based Sequence Transduction Neural Networks (Transformer) (US 10,452,978) - The foundational Transformer patent. Vaswani is first inventor. Introduces self-attention as the primary mechanism for sequence transduction, replacing recurrence and convolution. The 2017 architecture underpinning BERT, GPT, T5, PaLM, Gemini, and every modern large language model.
- Multi-Task Multi-Modal Machine Learning System (MultiModel) (US 20200089755) - A single neural network trained simultaneously on multiple tasks across multiple modalities (image, audio, text). Modality nets feed a shared body of convolutional, attention, and sparsely-gated mixture-of-experts blocks. The unified-architecture ancestor of MUM and Gemini.
- Attention-Based Image Generation Neural Networks (US 12,142,034) - Attention-based image generation. Multimodal application of the attention mechanism. Underpins image-search and Lens-style multimodal retrieval. Vaswani is co-inventor; canonical article lives in the Shazeer section.
- Local Self-Attention Based Computer Vision Neural Networks (HaloNet) (US 20210390410) - Vaswani is first inventor. Replaces convolution with blocked local self-attention plus overlapping haloed neighborhoods. Scales attention from quadratic to near-linear cost in sequence length. The scaling primitive behind sliding-window attention, Swin Transformer, and long-context Transformer ranking.
- Fully Attentional Computer Vision (Stand-Alone Self-Attention) (US 20250292560) - Replaces convolution operations with stand-alone self-attention as the primary building block of a vision backbone. Per spatial location, attention learns content-dependent receptive fields. The Vision Transformer (ViT) lineage that underpins Lens, Image Search, and multimodal grounding.
- Fast Decoding in Sequence Models Using Discrete Latent Variables (Latent Transformer) (EP 3732627) - Two-stage decoding. First predict a short sequence of discrete latent variables (compressed plan) from the source. Then decode the actual output sequence conditioned on those latents, with most per-token computation now running in parallel. The mechanism that makes AI Overviews, SGE, and Gemini-served features tractable at search-scale latency.
Jakob Uszkoreit, Google Transformer Architect, Universal Transformer & Vision Transformer Patents
- Attention-Based Sequence Transduction Neural Networks (Transformer) (US 10,452,978) - The foundational Transformer patent. Uszkoreit is co-inventor with Vaswani, Shazeer, Parmar, Jones, Gomez, Kaiser, Polosukhin. Canonical article in the Shazeer section; also cross-listed in Vaswani.
- Decomposable Attention Model for Natural Language Inference (arXiv 1606.01933) - The pre-Transformer ATTEND-COMPARE-AGGREGATE model that proved attention alone could replace recurrence on a hard NLP task. Three months before "Attention Is All You Need," this paper showed 10x fewer parameters than recurrent baselines could match or beat them on SNLI. The structural ancestor of the Transformer.
- Universal Transformers (Adaptive-Depth Transformer with ACT) (US 10,740,433) - The Universal Transformer: weight-tied recurrent Transformer block with per-token Adaptive Computation Time halting. Easy tokens halt early; hard tokens recur more times. Combines parallel attention with the inductive biases of recurrence. Precursor to dynamic-compute models, mixture-of-depths, and adaptive-cost LLM inference.
- Multi-Task Multi-Modal Machine Learning System (MultiModel) (US 20200089755) - Single neural network trained simultaneously on multiple tasks across multiple modalities. Uszkoreit is co-inventor. Canonical article in the Vaswani section.
- Vision Transformer (ViT): An Image Is Worth 16x16 Words (arXiv 2010.11929) - Pure Transformer applied to image classification by tokenizing the image as a sequence of fixed-size patches. No convolution. Pre-trained on web-scale data, ViT matches or beats state-of-the-art CNNs at scale. Foundation of modern visual search (Lens, Image Pack), multimodal grounding (CLIP, PaLI), and Gemini Vision.
- Attention-Based Image Generation Neural Networks (Image Transformer) (US 12,142,034) - Attention-based image generation. Uszkoreit is co-inventor. Canonical article in the Shazeer section; also cross-listed in Vaswani.
Simon Tong, Google Click-Driven Ranking, Population Signals & Anti-Spam Patents
- Improving Search Ranking Using Related Queries (US 7,505,964) - For each candidate document, aggregate ranking signal across a query AND its related-query cluster. Documents that perform well across coherent related queries get boosted; narrow single-query matches do not. Tong + Pearson + Sergey Brin. The structural ancestor of modern topical-authority signal.
- Improving Search Ranking Using Population Information (US 7,454,417) - Documents are ranked using population-information signal: which user populations (geographic, demographic, behavioral cohorts) engage well with the document for a given query. Audience-fit becomes a ranking signal. Tong + Pearson.
- Country Biasing of Search Results (US 20040254932) - Search results are biased toward the user-inferred preferred country. Country is inferred from IP, Google domain, language, prior preferences; document country inferred from ccTLD, hosting, language, content cues. Strong bias for navigational/local queries; weaker for global topical.
- Modifying Search Result Ranking Based on Implicit User Feedback (Navboost) (US 8,661,029) - The Navboost patent. Tong is co-inventor with Kim, Shazeer, Diligenti. Canonical article in the Kim section.
- Click-Through Rate as a Ranking Factor (US 10,229,166) - CTR-as-ranking-factor patent. Tong is co-inventor. Canonical article in the 65 Google Patents section (pat-55) and in Kim.
- Large-Scale Machine Learning Systems and Methods for Ranking (US 7,222,127) - The pre-Transformer-era distributed learned-ranker infrastructure. Trains a ranking model on billions of (query, document, label) tuples with distributed gradient-based learning across hundreds of features. Tong + Bem + Harik + Levenberg + Shazeer.
- Ranking Documents Based on Large Data Sets (continuation) (US 9,116,976) - Large-data-set ranking continuation. Tong is co-inventor. Canonical article in the Shazeer section.
- Historical Data (Information Retrieval Based on Historical Data) (US 7,346,839) - The Historical Data patent. Tong is co-inventor. Canonical article in the 65 Google Patents section (pat-51).
- Permitting Users to Remove Documents from Search Results (US 8,417,697) - User block-lists feed both personal-result filtering AND an aggregate crowd-sourced quality signal that affects ranking for other users. Tong + Ghemawat + Piscitello + Matt Cutts. Structural ancestor of Chrome Personal Blocklist (2011) and disavow-style signals.
Michelangelo Diligenti, Google Click-Data Quality & Implicit Feedback Patents
- Detecting Click Spam (US 8,694,374) - Builds an attribute model for normal network objects (users, sessions, IPs, cookies). Identifies objects whose attributes deviate from the model. Feeds deviance value back into the ranking algorithm to downweight anomalous clicks. Protects downstream ranking patents from polluted click signal.
- Click Model That Accounts for User Intent (US 20120143789) - Classifies query intent (navigational, informational, transactional, ambiguous, exploratory) and applies intent-conditional click weights. A click means different things for different intent classes; per-intent click models give cleaner relevance signal.
- Modifying Search Result Ranking Based on Implicit User Feedback (Navboost) (US 8,661,029) - The Navboost patent. Diligenti is co-inventor with Kim, Tong, Shazeer. Canonical article in the Kim section.
- Modifying Search Result Ranking Based on Implicit Feedback and a Model of Presentation Bias (US 8,938,463) - Click signal corrected for presentation-position bias. Diligenti is co-inventor. Canonical article in the Kim section.
- Click-Through Rate as a Ranking Factor (US 10,229,166) - CTR-as-ranking-factor patent. Diligenti is co-inventor. Canonical article in the 65 Google Patents section (pat-55).
Jaime Teevan, Microsoft Personalized Search & Search-Task Patents
- Personalized Search Results and Information Access (US 7,693,818) - Profile-vector re-ranking on top of a global candidate set. The foundational Bing personalization patent. Teevan + Dumais + Horvitz.
- Personalized Navigation Using a Search Engine (US 8,825,649) - Personalized navigation. Teevan is co-inventor. Canonical article in the Dumais section.
- Search Task Identification (US 8,326,824) - Cross-session query clustering identifies that a sequence of queries belongs to a single multi-step task. Per-task ranking conditions on task stage; serving the whole task journey beats single-query optimization.
- Re-Finding vs New-Finding Query Classification (US 8,756,219) - Classifies queries as re-finding (return to previously-clicked content) vs new-finding. Stabilizes prior clicked results for return visits so users can re-locate content reliably. Roughly 40% of queries are re-finding queries.
- Microtask Search Decomposition (US 10,062,103) - Decomposes high-level goals (plan a trip, evaluate a product, complete a workflow) into resumable micro-steps with step-conditional ranking and cross-session progress tracking. The microproductivity research direction underpinning M365 Copilot.
Eric Horvitz, Microsoft Decision-Theoretic Search & Attention-Aware Information Delivery Patents
- Decision-Theoretic Ranking Under Uncertainty (US 7,124,129) - Relevance is modeled as a probability distribution rather than a deterministic score. The ranker selects results that maximize EXPECTED UTILITY across the uncertain relevance distribution. Multi-signal sufficiency beats single-signal maxing.
- Cost-Benefit Reasoning in Search-Result Presentation (US 7,194,464) - Inline vs link decision is priced as expected benefit minus attention cost minus risk of being wrong. The mechanical basis for featured snippets, knowledge panels, and AI Overviews. Definitive, low-risk answers earn inline placement.
- Attention-Aware Proactive Information Delivery (US 6,934,917) - Context + attention state + need prediction drives zero-query information surfaces (Discover, Now cards, AI assistant suggestions). Discoverability extends beyond the search box into proactive feeds.
Ryen White, Microsoft Search Behavior & Search-Trails Patents
- Search Trails / Successful-Path Mining (US 7,904,440) - Aggregates cross-user query+click trails. Ranks by which destinations end successful trails. Being the end-of-journey page beats catching the first click. The structural ancestor of journey-aware ranking and AI-Overview citation selection.
- Exploratory vs Lookup Search Classification (US 8,041,711) - Classifies query phase as exploratory, lookup, or mixed. Serves phase-tuned ranking and result formats. Exploratory queries get overview/comparison content; lookup queries get definitive answers.
- Cross-Session Task Continuation (US 9,116,996) - Stitches queries across sessions and devices into one continuing task with stage tracking. Content covering the whole task arc (research, comparison, decision, post-purchase) wins over single-stage sources.
Azalia Mirhoseini, Google ML Infrastructure & Chip-Design Patents
- RL-Based Chip Floorplan Placement (US 11,475,278) - RL graph-placement policy lays out integrated-circuit floorplans in hours that previously took human engineers weeks. Feeds back into TPU generations. The compute-economics lever that lets Google afford Transformer-scale ranking per query.
- Device Placement for Distributed ML Training (arXiv 1706.04972 (Google patent family active)) - RL controller learns how to split large neural networks across GPUs/TPUs to minimize training step time. The infrastructure layer that made training BERT, T5, PaLM, and Gemini feasible at production scale.
- Mixture of Experts Neural Networks (US 11,769,055) - The sparsely-gated Mixture-of-Experts patent. Mirhoseini is co-inventor with Shazeer, Dean, Hinton, Le. Canonical article in the Shazeer section.
Arnold Overwijk, Microsoft Site Topical Authority Patent
- Extracting Key Phrase Candidates and Producing Site-Topical Authority Ranking (US 20210004416) - Per (site, topic), topical authority A(s,t) = SR(s) * C(s) * T(s,t)^2. SR = site rank, C = click signal, T = per-topic page share squared (strongly rewards depth). ML-trained key-phrase extraction discovers each site's topics. The most explicit topical-authority formula in the modern literature.
Michael Procopio, Google Per-Author Per-Topic Authority Patent
- Per-Author, Per-Topic Authority Signatures (US 8,458,196) - Per (author, topic), authority signature = sum over the author's documents of (authorship % * topic weight). Quantifies "this person is an authority on this topic." Recurring focused contribution concentrates the signature; tangential one-off pieces dilute it. Structural basis for the E-A-T Expertise signal at the author level.
Ori Allon, Google Orion Query-Refinement Patent
- Refining Search Queries (Orion) (US 8,392,443) - Generates ranked query refinements by mining phrase patterns from documents that satisfy the parent query. Corpus-driven (content-side) refinement, complementing log-driven (behavior-side) refinement. Drives "Searches related to," extended snippets, and the People-Also-Ask lineage.
Corin Anderson, Google Reasonable Surfer (Foundational) Patent
- Reasonable Surfer (Foundational): Ranking Documents by User Behavior & Feature Data (US 7,716,225) - The foundational Reasonable Surfer patent. Anderson is first inventor with Jeff Dean and Alexis Battle. Per (source page, link) pair, click probability varies with link features (anchor text, font size, position, surrounding context, link type). PageRank flow weighted by per-link probability.
- Reasonable Surfer Model (2012 continuation) (US 8,117,209) - The 2012 continuation in the Reasonable Surfer family. Anderson co-inventor. Canonical article in the 65 Google Patents section (pat-45).
Ashok Popat, Google OCR & Image-Text Indexing Patents
- Translation-Inspired OCR (US 8,675,012 family (Popat et al., research-grade SMT-OCR)) - Reframes OCR as statistical machine translation: image features map to characters via a channel model; a language model scores candidate output fluency; joint decoding maximizes channel times language likelihood. Multi-language coverage scales by swapping models. Drives image-text indexing across Google Search, Books, Lens, and Drive.
Why this organization
Patent law clusters intellectual property by inventor and assignee, not by topic. Grouping by inventor recovers the narrative arc: Amit Singhal's decade of ranking work, Pandu Nayak's BERT-era query handling, Krishna Bharat's news ranking lineage. Each track reads like a research biography written in claim language.