Since January 2024, Google and DeepMind have published patent applications and granted patents covering 4,651 inventions in search, language processing, AI, advertising and computer vision. Most of them have nothing to do with the search engine. Of the 339 that an AI model rated as directly related to search, 93 have a write-up in at least one of three sources that track Google patents for SEO. That is roughly one in four. AI answers are written up almost half the time, Google Maps almost never: 3 write-ups for 66 inventions.
I put all 4,651 inventions into one database. Each has a title, an abstract (the applicant’s own summary), dates, a topic area, a score for how closely it relates to search, and a flag showing whether anyone has already written about it. You can download the database under a CC BY 4.0 license. I count an invention as a patent family, which means all publications of the same invention in different countries. An AI model scored the connection to search, and I checked its scores on a sample with a second model. The method is at the end.
Key numbers from the Google patents database
The database covers 4,651 inventions and 11,632 publications: applications, granted patents and their counterparts in other countries, published between January 1, 2024 and September 17, 2026.
- 1,088 inventions may relate to search and visibility. An AI model scored every invention on a 0–3 scale, and these 1,088 got a 2 or a 3. The 339 with a 3 are directly related to the search engine and AI answers. A check on a sample suggests the real number of search-related inventions is lower, roughly 600.
- One in four inventions directly related to search has a write-up. I checked three sources that write up Google patents for SEO: 93 of the 339 have a write-up. The weaker the connection to search, the less often anyone wrote about an invention: 8.1% at a score of 2, under 2% at scores of 1 and 0. Across the whole database, 177 inventions have a write-up.
- 97 new filings with the top score have no write-up in any of the three sources. Their priority date is 2023 or later. The priority date is the date of the first filing for an invention. These 97 are where I would start reading.
- One in three inventions from 2024 mentions a language model. 34.9% of inventions first filed in 2024 mention a language model or a generative model in the title or abstract. In 2020 it was 1.5%.
- 58.8% of inventions have a granted patent in the database, in at least one country. The rest only have applications, which may still be rejected or abandoned.
A patent does not tell you how the search engine works. It tells you what solution Google disclosed and what protection it applied for. So the database shows what Google files for protection, not a list of ranking factors.
What Google patents in search
One in three of the 1,088 search-related inventions deals with AI answers and information retrieval: 363. Next come ads (155), personalization and recommendations (145), images and video (134), and Maps (93).
The model assigned an area and a score to each invention based on its title and abstract, so treat the breakdown as a map, not a verdict. Five things stand out:
- AI answers are growing. Among inventions with a priority date of 2023 or later, information retrieval and AI answers make up 40.5% of the pool, compared with 27.8% among older ones. Almost two thirds of the new inventions in this area (65.3%) mention a language model or a generative model in the title or abstract.
- Maps are mostly old inventions. Half of the 93 have a priority date of 2018 or earlier, and Google filed only 16 of them in 2023 or later. Maps account for 3.4% of the pool from 2023 on and 12.6% of the older pool.
- Ads are campaign technology. Of the 155 ad inventions, the model rated only 2 as directly related to search results. The rest cover topics such as conversion attribution and campaign optimization.
- Crawling and indexing: 16 inventions. None of them have a priority date of 2023 or later.
- Links: 2 inventions in the whole pool. Both are new, and neither has a write-up. I describe them below.
What has already been written up, and where
Of the 1,088 search-related inventions, 154 have a public write-up, and most of these, 81, deal with AI answers and information retrieval. I checked three sources that systematically write up Google patents for SEO:
| Source | Inventions written up (of 1,088) | What it is |
|---|---|---|
| Olaf Kopp | 114 | patent write-ups in English, some behind a paywall; matched by title, so the count is approximate |
| Semgence, public patent catalog | 65 | 450 publications, including 371 from Google; catalog as of August 17, 2026 |
| Rafał Borowiec | 26 | analyses in Polish |
Bill Slawski’s archive ends in 2022, and I found none of the inventions from this database in it. Each author picks patents by their own criteria, so the coverage shows what these three sources focus on, not what matters to Google.
A patent number ends with a kind code. In the US, A1 means a published application, and B1 and B2 mean a granted patent. The WO prefix marks an international application.
Eight inventions have a write-up in all three sources:
| Patent | First filed | What it is about, according to the abstract |
|---|---|---|
| US12326889B2 “Contextual estimation of link information gain” | 2018 | information gain: an estimate of how much new information a document adds beyond what the user has already seen |
| US12579133B2 “Generating query variants using a trained generative model” | 2017 | a generative model creates variants of a query |
| US12248529B2 “Multi source extraction and scoring of short query answers” | 2022 | a model that scores text passages picks a short answer to show in a box |
| US20240289407A1 “Search with stateful chat” | 2023 | a chat that remembers the context of the session, added to regular search |
| WO2024229375A1 “Query expansion by prompting large language models” | 2023 | query expansion by a language model |
| US20240362093A1 “Query response using a custom corpus” | 2023 | a language model answers based on documents that a third-party application can access |
| US12670323B2 “Golden prompt generation based on authoritative publications” | 2023 | model prompts built from authoritative publications on a topic |
| US20260064780A1 “Dynamic attribution and/or modification of responsive content…” | 2024 | when to modify parts of an AI answer built from documents in the search results |
Does no write-up mean not interesting?
Partly, yes. The stronger the connection to search, the more write-ups: 27.4% of inventions with a score of 3 have one, 8.1% with a score of 2, 1.5% with a score of 1 and 0.3% with a score of 0. Authors write about what matters for SEO and mostly skip ad technology, personalization and general machine learning. That is a sensible choice.
Two gaps cannot be explained that way. The first is Maps: of the 66 inventions that the model rated as directly related to search, 3 have a write-up, while 52 of the 116 on AI answers do. Age explains part of it, because half of the Maps inventions have a priority date of 2018 or earlier. But the 11 new Maps filings with the top score and no write-up include AI-generated business profiles.
The second gap is time. Among inventions on AI answers, 50% of those first published in 2024 have a write-up, compared with 29.2% of those published in 2026. That looks more like a lack of time than a lack of substance.
Where to start reading
I would start with the 97 new filings with the model’s top score that none of the three sources have written up. Most of them, 43, deal with AI answers and information retrieval, 23 with images and video, 11 with Maps and 9 with quality and trust.
I went through their titles and abstracts and picked the most interesting ones. I describe them based on the abstracts, the summaries written by the applicant. The scope of protection is defined by the claims, the numbered statements at the end of a patent, and I did not analyze those here.
Maps and business profiles
- AI-generated business profiles. US20250054045A1 “Artificial Intelligence Generated Business Profiles”, an application with a priority date of August 10, 2023. The system creates a business profile in response to a query, from data about the business and data about the user who is asking. Two people could therefore see different profiles of the same business.
- Place attributes from web pages and queries. US20260267829A1 “Automatic Generation of Schemas and Attributes for Geographic Locations”, an application with a priority date of March 7, 2025. Machine learning models derive schemas from web page content and queries, then generate attributes for content categories. According to the title, it is about geographic locations.
- Place data as context for a language model. WO2024233832A1 “Augmenting large language models understanding of web information with factual and subjective location based data”, an application with a priority date of May 9, 2023, also filed with the European Patent Office (EPO) and in China. The model answers questions about places based on structured location data. The title points to two types of that data: factual and subjective.
If you manage business listings, compare these three applications with what Google itself says about local ranking.
AI answers and search results
- Links in a model’s answer lead to follow-up questions. US20260178846A1 “Using llms to automatically create follow-up links in llm response”, an application with a priority date of December 19, 2024. The model writes its answer in a markup language such as HTML and adds a set number of links on key phrases. Each link leads to a follow-up query, and the highlighted phrase is clickable.
- AI badges in results. US20240378256A1 “Artificial Intelligence Generated Badges for Search”, an application with a priority date of May 9, 2023, also filed in Japan. The system derives attributes of an entity from information on the web and turns them into badges. According to the abstract, the badges may affect ranking and label results in the interface.
- Answer format by intent. US20260252621A1 “Identifying Content Formats Based On Search Query Intent”, an application with a priority date of November 28, 2023, also filed with the EPO. An intent indicator determines whether the user wants an image, a video, text or audio.
- AI-written video descriptions. US12566796B1 “Query-dependent generative descriptions for videos provided via a search result”, a granted patent with a priority date of October 4, 2024. A model writes a description of a video for the search result based on the query and the video content. The description appears in the results once its consistency score against the input data reaches a threshold.
- A model trained on the difference between answers with and without search results. US12437016B2 “Fine-tuning large language model(s) using reinforcement learning with search engine feedback”, a granted patent with a priority date of December 7, 2023. That comparison produces a signal for training a reward model, which then fine-tunes the language model.
- Pages generated on demand. US20250094521A1 “Generative navigational corpus”, an application with a priority date of September 19, 2023. A large model creates a web page from source content, matched to the user’s intent.
Ads and links
- Ads in generative formats. US20260003902A1 “Generative ai search formats”, an application with a priority date of July 1, 2024. One AI model creates categories related to the query, a second one writes a description for each, and sponsored content and links to external pages appear alongside.
- A model checks whether a link leads where it promises. US20260099559A1 “Systems and Methods for Link Validation with Machine Learning”, an application with a priority date of October 9, 2024. The system takes a screenshot of the target page and asks a multimodal model whether it matches the link text. It also detects mismatches in meaning.
A pre-ChatGPT invention, also without a write-up
- A model’s answer based on search engine results, filed before ChatGPT. US12333436B2 “Augmenting machine learning language models using search engine results”, DeepMind, priority date January 31, 2022. The system turns a question into a search engine query and builds the input for a language model from passages of the documents it finds. The family has two US patents and publications at the EPO and in the UK, Korea, China and Germany.
You can pull the full list of 97 filings from the database with one filter, which I give in the section on using the data.
Odd finds in the database
The oldest invention is more than 20 years old. US12008727B2 “Digital mapping system” has a priority date of March 23, 2004, less than a year before Google Maps launched in February 2005. The patent was published on June 11, 2024. It describes an info window that opens on the map when you select a marker. The database has 33 inventions first filed before 2010 and 282 before 2015.
In 2004, Google filed a patent on fetching results for questions the user had not asked yet. US11971897B2 “Predictive information retrieval”, with a priority date of December 3, 2004, describes a system that predicts the next queries after a query is submitted and fetches information for them in advance. The patent was published on April 30, 2024.
Nihil novi sub sole.
Google patents don’t say “Google.” In the titles and abstracts of the 4,651 inventions, the words Google, YouTube, Gemini, Chrome, Android and SEO do not appear once. A patent describes a solution, not a product, so without knowing the language of patents, they are hard to find. You will find a glossary below, after the section on reading patents.
Google files most widely for inventions about the models themselves. The record in the database, 30 publications at nine patent offices, belongs to audio generation with neural networks (US12322380B2). Among search-related inventions, the most publications, 25, belong to a language model that produces an intermediate text analysis before it answers, using tools such as APIs (US12481834B2).
DeepMind has 440 inventions in the database, but the AI model linked only 24 of them to search. The rest are mostly about images, robotics and machine learning.
The AI model makes mistakes too. It rated a patent on choosing video thumbnails based on viewer activity as directly related to search. The second model gave it a 1 on the 0–3 scale. That is why I treat the scores in the database as a reading filter.
Language models in Google patents
The share of inventions that mention a language model or a generative model in the title or abstract grew from 1.5% in 2020 to 34.9% in 2024. I assign each invention to a year by its priority date, because an application usually becomes public after 18 months, and a patent is often granted several years later.
The share jumped from 8.4% to 29.1% in 2023, the first full year after ChatGPT launched on November 30, 2022. The direction will surprise no one. What is more interesting is that language models also made it into inventions about images and video: 35.5% of the new ones mention them, compared with 4.2% of the older ones.
The flag is a search for terms such as “generative model,” “large language model” and “LLM” in the title and abstract, not a score from the AI model.
What the claims of nine AI-answer filings say
In nine newer filings on AI answers, I read the claims, the numbered statements that define the scope of protection. I picked them from the 97 without a write-up, and I describe claim 1 and selected dependent claims. Google Patents marks three of them as abandoned. I give the statuses as shown in Google Patents on October 27, 2026, and Google Patents itself labels the status as an assumption, not a legal conclusion.
- Sub-queries. WO2025080561A1 “Utilizing generative model in responding to multifaceted queries”, priority date October 10, 2023. A generative model breaks a question into sub-queries, the system selects some of them, retrieves search results for each one and builds the answer from those results. The international application has the status “Ceased” in Google Patents, but applications from this family at the EPO, in China and in Korea are pending. In its documentation, Google calls the technique of running multiple related searches query fan-out, but that does not prove it works the way this application describes.
- Relevant and diverse passages. US20260072965A1 “Controlled content diversity in retrieval for generative search”, September 6, 2024, pending. For a query and related queries, the system selects passages from the top-ranked pages by relevance and diversity, and a language model writes the answer from them. Claim 5 requires passages shorter than 500 characters.
- Only the matching part of a page. US20260187163A1 “Selecting a portion of content for a generative model”, December 31, 2024, pending. The system compares parts of a page with a string of words that may include the user’s query and passes to the model only the parts that meet a threshold condition. According to claim 3, a skipped part may contain an instruction for the model.
- Source quality when choosing an answer. WO2025170579A1 “Generative model tuning and inference utilizing quality signals”, February 7, 2024, pending. The model generates several candidate answers, and the system picks one based on the quality scores of the pages associated with those candidates, meaning the scores of those pages as search results.
- Passage reliability. US20260211889A1 “System for controlling response reliability”, January 23, 2025, abandoned. The model receives passages together with their reliability scores and uses those scores to weight the passages in its attention mechanism. The score may depend on where a passage comes from.
- Source scores against hallucinations. US20260134225A1 “Tracking identified sources for attributions to reduce hallucinations”, November 11, 2024, pending. Each candidate answer comes with a list of sources and their scores, and the system picks a candidate based on those scores, a threshold and a source the user specified. Claim 9 lists news services, social media, data aggregators and scientific publications.
- Content opted out by its owner. US20260134028A1 “Modification of responsive content that is generated using generative model(s) and that includes opt-out content”, November 11, 2024, pending, also at the EPO. If someone has opted out of having their data used to train a model or to generate answers, and part of an answer matches that content, the system modifies that part.
- Page signals inside a language model. US20250103662A1 “Unifying transformers with link based ranking”, September 21, 2023, abandoned. The language model adjusts its attention mechanism with signals from the page the text comes from. Claim 2 lists signals such as the author, the page’s ranking, traffic statistics, creation and modification dates and an engagement metric.
- Social media posts. US20250156488A1 “Determination of user post items related to search content using creator context data”, November 15, 2023, abandoned. The system matches user posts to content, for example a news article, taking into account data about the authors: name, bio, website address and location.
If you read patents as a list of ranking factors, applications 5 and 8 would top it: source reliability in AI answers, and the traffic and author of a page. Google Patents marks both as abandoned. An abandoned status only describes the patent proceedings. It says nothing about whether Google uses the technology, and a pending application does not prove that either.
From these nine, I would take three hypotheses to test: whether short, self-contained passages make it into AI answers more often, whether a page appears in answers to related questions and not just to the main one, and which sources AI answers cite most often in your industry. You can test each one on your own data, for example by tracking which pages AI answers cite for the queries that matter to you.
How to read a Google patent
What a patent protects is decided by the claims, not by the title or the abstract. I read a patent in four steps.
- The code at the end of the number. In the US, A1 means a published application, and B1 and B2 a granted patent. At the European Patent Office, A1 is an application published with a search report, and B1 is the specification of a granted patent. A number with the WO prefix is an international application.
- Priority date and status. Later filings of the same invention in other countries share its priority date, and so do continuations, which are later US applications based on an earlier one. That is why a patent published in 2026 may have been first filed in 2011. The status in Google Patents is an assumption, not a legal conclusion. Check the binding status in the patent office register.
- Independent claims, the ones that do not refer to other claims. Dependent claims describe variants.
- The description and drawings, to understand the terms used in the claims. The description may present many variants, and none of them has to be implemented.
Caution: a patent does not prove implementation, signal weight, or that a solution is used in your market. Words also mean something different in patents: “entity” or “document” can be any object, not an entity or a page in the SEO sense. Check every conclusion from a patent against Google Search Central documentation or your own data. Google’s own overview of crawling, indexing and serving results is in How Search works.
How Google names SEO concepts in patents
Patents describe the search engine in their own language. A page is usually a “resource” or a “document,” a passage is a “portion,” and an AI answer is a “response.” If you search patents with SEO vocabulary, you will miss most of them.
| In SEO you say | In Google patents | Titles and abstracts with the word (of 4,651) |
|---|---|---|
| page | resource / document | 233 / 225 |
| web page / website | 43 / 21 | |
| part of a page | portion | 358 |
| snippet / passage | 27 / 10 | |
| AI answer | response | 790 |
| language model | language model, LLM / generative model | 354 / 240 |
| search result | search result | 125 |
| click | user interaction, user engagement / user selection | 77 / 43 |
| click / dwell time | 6 / 0 | |
| entity | entity / knowledge graph | 170 / 10 |
| author | creator / author | 13 / 1 |
| link | link | 53 |
| crawl / spam | crawl / spam | 3 / 1 |
The counts cover titles and abstracts only. Claims and descriptions may use these words more often, and 292 inventions have no abstract in the database. Still, the difference in scale is clear: search for “click” and you find 6 inventions; search for “user interaction” or “user engagement” and you find 77. The same goes for crawling. The “Crawling and indexing” area has 16 inventions, but the word “crawl” appears in the title or abstract of only 3 inventions in the whole database, because the model assigned areas by content, not by keyword.
How to use the database
You can download the database as one file: google-patents-2024-2026.zip. It contains a CSV with the 4,651 inventions, the validation sample, the SQL query and the scoring instructions for the model. Each row has the number, a link to Google Patents, the title, the abstract, dates, whether a patent was granted, the score and area assigned by the model, the language model flag and whether the invention has a public write-up.
Four filters I would start with:
- New, no write-up: model_score_0_3 = 3, priority_date from 20230101, written_up = no. The filter returns 97 filings.
- Local SEO: model_area = “Maps and local results” and a score of 2 or 3. The database has 93 such inventions, 66 of them with a score of 3, and someone has already written up 4 of the 93.
- Already written up: written_up = yes. You get 177 inventions to compare with what has been written about them.
- Patent vocabulary: search the title and abstract columns for words from the glossary above, for example “portion” together with “generative.”
You can read individual patents for free in Google Patents. For analysis at scale, use the same data in BigQuery: my query processed 257 GB, and the first 1 TiB of queries each month is free for a billing account, as long as other queries that month have not used it up. The source data, Google Patents Public Data by IFI CLAIMS Patent Services and Google, is licensed under CC BY 4.0, and so is my database: you can use it as long as you credit the source. I will update the database in April 2027.
Watch out: the data in BigQuery lags behind Google Patents and is refreshed from time to time. In this database it ends on September 17, 2026, and the same query run later will return a different set.
Patents give you hypotheses, but you set your budget and priorities based on traffic and conversions. Checking how a brand shows up in AI answers, and what to change, is part of my AI search optimization work.
How I built the database
I pulled the data from Google Patents Public Data with a single SQL query, and an AI model scored each invention for its connection to search.
| Parameter | Value |
|---|---|
| Source | Google Patents Public Data, the public dataset of patent publications from IFI CLAIMS Patent Services and Google |
| Applicants | Google LLC and DeepMind |
| Window | publications from January 1, 2024 to September 17, 2026 |
| CPC classes | at least one of: G06F16 (information retrieval), G06F40 (natural language processing), G06N (machine learning), G06Q30 (commerce and advertising), G06V (image recognition) |
| Unit | patent family |
| Year | year of the priority date of the publication that represents the family |
| Result | 4,651 families, 11,632 publications |
Title, abstract and priority date come from one publication per family, in this order of preference: US granted patent, US application, WO, EP. CPC, the Cooperative Patent Classification, is a topic classification assigned by patent offices, so this filter does not depend on the model.
Model scores and the validation sample
The AI model scored each family based on the title and up to 700 characters of the abstract, on a 0–3 scale:
- 3: directly about the search engine and AI answers, for example ranking, indexing, source selection, citations, snippets, Maps, Shopping, Discover, News: 339 families;
- 2: a mechanism that may affect visibility, for example query understanding, entities, fact-checking, recommendations, ads in search results: 749;
- 1: general machine learning and language processing: 932;
- 0: images, hardware, speech, robotics and other areas: 2,631.
The model also assigned a topic area. Its one-sentence notes on the impact on visibility are untested hypotheses, so they are not in the database.
I checked the scores on a sample: 15 random families from each score, 60 in total. A second language model from a different developer scored them using the same criteria, based on the title and the full abstract, in random order and without seeing the first score.
| First model’s score | Second score: 3 | 2 | 1 | 0 |
|---|---|---|---|---|
| 3 (15 families) | 10 | 3 | 2 | 0 |
| 2 (15 families) | 3 | 3 | 9 | 0 |
| 1 (15 families) | 0 | 0 | 13 | 2 |
| 0 (15 families) | 0 | 0 | 3 | 12 |
The scores agree in 49 of 60 cases if you only count the split between search-related (2–3) and the rest (0–1). Most disagreements concern score 2: 9 of 15 such families got a 1 from the second model. Projecting the sample results onto the whole set gives 593, roughly 600 search-related families instead of 1,088. This estimate assumes that the groups the first model scored 0 and 1 contain no relevant families, and a sample of 15 per group cannot rule out a true share of around 20%.
Limitations
- Model scores: the model assigned scores and areas without reading the claims. 292 families have no abstract, so the model scored them by title alone. None of the scores in the database is an expert assessment.
- Class scope: the database covers only classes G06F16, G06F40, G06N, G06Q30 and G06V. Of the 74 Google families from 2024–2026 in the Semgence catalog, 4 fall outside this filter.
- Publication window: a family enters the database when a publication that meets the filters appeared in 2024 or later, so older and newer years had different chances of getting into the set. The comparison of years therefore describes the database, not Google’s whole portfolio. Most applications from 2025 and 2026 are not public yet: the database has 52 inventions with a priority date in 2025 and one in 2026.
- Language model flag: it is a term search, so it also catches mentions in the background of an invention and misses patents that use other words.
- Write-ups in the sources: matching by number and title may have missed some write-ups or counted titles whose full write-up is paid. There are also paid communities with patent analyses, and I did not use their content.
- Status: I checked the status only for the nine filings whose claims I read. The granted_patent column only says whether the database has a document with a kind code starting with B.
If you know of a write-up of a patent from this database that I missed, let me know and I will update the data.
FAQ
Does a Google patent mean Google uses that mechanism in search?
No. A patent shows a disclosed solution and the protection applied for, but it does not prove implementation or signal weight. The database includes inventions first filed as early as 2004, and Google Patents marks three of the nine newer AI-answer filings whose claims I read as abandoned. An abandoned application does not prove that Google does not use the solution either.
How many Google patents relate to SEO?
Several hundred. According to an AI model’s scores, 1,088 of the 4,651 inventions published since 2024 may relate to search and visibility, including 339 directly. A check on a sample suggests the real number is lower, roughly 600. About one in four of the 339 inventions directly related to search has a public write-up.
Where can I find Google patents, and is it free?
In Google Patents, for free. That is where you read individual patents. For analysis at scale, use BigQuery, where the first 1 TiB of queries each month is free. My query processed 257 GB. You can download the ready-made database of 4,651 inventions from 2024–2026 from this post.
What is the difference between a patent application and a granted patent?
An application is a request that is still waiting for the patent office’s decision, while a granted patent provides protection. In the US, the code A1 marks a published application, and B1 and B2 a granted patent. In the database, 58.8% of inventions have a granted patent in at least one country.
What is a patent family?
It is all the filings of one invention, linked by a shared priority, often from several patent offices. In the database, 4,651 families account for 11,632 publications, which is why I count families, not publication numbers: counting individual publications would inflate the total.
How long does a patent last?
Usually 20 years from the relevant filing date, both in the US and for a European patent. Protection may end earlier, and the law provides for exceptions and extensions. You can check the term of a specific patent in the register.
Leave a comment