What Google's Patents Reveal About How AI Search Chooses Sources

Google's patents describe how AI Overviews choose sources as two passes over search results. The first retrieves ranked pages for the question and the related searches an AI model writes, and the second keeps a few for the answer, using measures such as ranking position, clicks, trust, freshness, and location.

Quoted alone, these filings read like rules for Google AI Overviews, a feature with over 2.5 billion monthly active users. Google confirms retrieval through its core ranking systems and query fan-out, while none of the filings shows which selection measures a Google product runs.

For a SaaS, fintech, or AI company, that gap decides where the website budget goes first. A patented measure Google has not confirmed can pull that budget toward the wrong fix, while a key page still lacks eligibility, a clear answer, or original information. This guide shows which steps Google confirms and the website changes they justify.

Key takeaways

• Google's summaries patent describes retrieval, then selection: documents are retrieved for a search, a second pass selects material for the summary, and supporting links can be found separately.
• The detailed measures quoted in posts sit in the patent's description, while claim 15 names three broad classes of measures.
• Google confirms retrieval through core ranking and query fan-out, and it says it ranks links inside AI answers, but it does not confirm the patent's second-pass measures.
• Information gain comes from a separate patent granted in 2022 that scores a document against what one person has already been shown, and Google has not named it as a signal.
• The work starts with what Google documents: eligibility, a useful page for each question prospects ask, the click, visible authorship, real updates, market versions, and original information.

What is AI Overview source selection, and what do Google's patents prove?

AI Overview source selection covers two jobs: choosing the web pages that inform an answer written by large language models (LLMs), and choosing the pages that support its links. The summaries patent describes these as separate operations. A company's own pages are one source among many. Google says its generative AI features show "what's being said about products and services across the web", in blogs, videos, and forums.

A patent has two parts that matter here. The claims define the scope of protection Google sought and was granted, while the description lists options an engineer could build. In Google's generative summaries patent, the measures quoted in posts sit in the description. The claims name only three families of measures, in claim 15, which builds on a claim about confidence labels: "query-dependent", "query-independent", and "user-dependent".

The patent's first claim covers something else: revising an AI-generated summary after the person interacts with one of its linked documents, for example by viewing part of it. So the paragraph that circulates as a rule is, in legal terms, one option among several.

Neither part proves use. A patent records a method Google chose to protect. A granted patent has passed examination, while publishing an application only makes it public, whatever its examination status. So each step in this guide is checked against Google's documentation, its blog, and sworn testimony in the US antitrust case against Google.

Google's own AI Mode shows how easily a patent and a product get mixed. Asked on October 27, 2026, what Google's patents reveal about cited pages, it named the fan-out and thematic search filings. It skipped the selection measures and said pages "bloated with sales pitches are systematically bypassed", a rule none of the selection passages in these patents states.

The court in United States v. Google found that some evidence suggests AI Overviews have "reduced user interactions with organic web results". It also cited a Google document saying pages that appear as a corroborating link "get more clicks" than the same page among the blue links for the same query. Google made the same claim publicly in May 2024, without a figure.

Five dated changes from the past six months show how quickly the area moves:

• May 15, 2026: Google published a guide to optimizing websites for its generative AI features.
• May 26, 2026: Google was granted US12639300B2, a patent on answering multifaceted questions with model-written subqueries.
• May 27, 2026: Preferred Sources reached AI Overviews and AI Mode, highlighting websites a person has chosen.
• August 31, 2026: A Search Console setting for generative AI features reached all websites worldwide.
• October 15, 2026: Google announced more prominent inline links and plans to test a section at the top of some answers for the "one or two official places on the web" where a task gets done.

Six patent families sit behind this guide. Two of them, the summaries and information gain patents, describe choosing documents, though no source joins them into one pipeline:

Checking a patent screenshot. Search the quoted sentence on Google Patents, see whether it sits under Claims or Description, and read the kind code at the end of the number: B1 or B2 marks a granted US patent, and A1 a published application. A continuation is a later application built on an earlier application's disclosure, with no new matter, and it can claim something different; together the filings form a family.

How does Google AI Overview choose sources, stage by stage?

The five stages below organize patent proposals and Google's confirmed steps into one model, which is our synthesis. The flow shows the four that follow eligibility, and eligibility, fan-out, and ranked results are the three Google confirms most directly.

1. Eligibility and the decision to show an AI answer. A page can be a supporting link only if it is "indexed and eligible to be shown in Google Search with a snippet", Google's AI features documentation says. Since August 31, 2026, every website can also opt out of generative AI features in Google Search Console. Opting out ends traffic and impressions from those features, and the nosnippet and max-snippet controls limit what can be shown.

Next, Google decides whether an AI answer helps at all. US12118325B2 claims deciding this from the terms a query includes or omits. Liz Reid, who leads Google Search, testified that an AI Overview appears "whenever we think that it is both high quality information and a net add to the page".

So an AI Overview appears only for some queries. In March 2025, Google said AI Mode falls back to web results without "high confidence in helpfulness and quality".

2. Fan-out: one question becomes many searches. A model then widens the question. Google confirms this for AI Mode, which splits a question "into subtopics" and runs "a multitude of queries simultaneously", and Deep Search "can issue hundreds of searches". For AI Overviews, Google's documentation says they "may use" the same technique. Its 2026 guide lists fan-out among the techniques both features rely on.

Three filings describe how the extra search queries are written. The stateful chat application builds synthetic queries from the question and the session, such as earlier queries and the pages returned for them, and can also show them as suggested next questions. The multifaceted queries patent has a model write candidate subqueries, then keeps those related to the question and distinct from each other.

The family's published application says subqueries can be generated when an input is longer than a threshold, is submitted less often than a threshold frequency, or returns low-quality results. A detailed B2B software question is a good example, since it combines several needs at once.

The query variants patent adds a loop. A trained control model reads the results of a first variant and then decides whether to write more. Google describes a similar loop for AI Mode, which adjusts its plan "based on what it finds", and one claim of the multifaceted patent ties subqueries to server load. So the searches behind one question can differ from run to run.

Read more: Query fan-out

3. Ranked results for each search. Each search then returns ranked results from Google's index, as on a search engine results page (SERP), and here the patents and Google agree. The multifaceted patent takes "top ranked search result(s) for each subquery", and the summaries patent can take the top few results. Google says the AI Overviews model "is integrated with our core web ranking systems".

In the summaries patent, material for the summary comes from documents responsive to these searches, and corroborating pages can be retrieved separately. Google's eligibility guidance gives no required organic position. Yet the searches are wider than the words typed. Google says its models "identify more supporting web pages" while a response is being generated. The summaries family also describes linking a page found by a search on one of the answer's own sentences.

So a cited page need not appear on the first results page for the typed query. Google retrieves supporting pages through Google Search, including the related searches it writes, and says this gives its AI features "a wider and more diverse set" of links than classic search.

4. A second pass picks the pages the model reads. From that pool, the generative summaries patent describes a second pass that keeps a few documents. It uses query-dependent, query-independent, and user-dependent measures, each set out as a factor below. Read side by side, the stateful chat application, filed a year later, repeats the selection paragraph word for word.

Google says it ranks the links inside its AI answers: it is "continuing to enhance how we show and rank links in our generative AI Search experiences". The post does not name the inputs, so it does not confirm the patent's measures.

The pool can also widen when the first results are weak. In one implementation, documents for a related query enter the set only when the first results "are of low quality and/or not diverse relative to one another" and the query correlation meets a threshold. That correlation can reflect how often one device or account submits both queries close together, and separate options add documents for recent or implied queries.

We read this as an opening on thin B2B questions, where a page ranking for a closely related question can enter the set a summary is written from.

5. The answer is written and its links attached. The model then writes from the kept set, and the patent's own example uses content from five documents. Portions of the answer are linked to the pages that support them. Google has also introduced previews of perspectives from forums and social media, and suggestions for in-depth reading, in many AI responses.

A sixth filing, thematic search, concerns the results page rather than the answer. A model summarizes passages across results, groups them into themes, and can rank themes by "a number of distinct documents" that mention each, among other signals. Google has shipped AI-organized results pages for recipes and a Web Guide experiment, but no source ties either one to this patent.

What are the AI Overview ranking factors in Google's patents?

The summaries patent describes six groups of document-selection measures, and information gain is a separate invention that scores new documents against what one person was already shown. No selection passage gives a weight, so each measure is a patent option, never a formula. Each one notes what Google confirms for Search and for its AI answers.

1. Ranking position for the question and its fan-out searches. The summaries patent lists a document's "positional ranking" for the query as one possible measure for selecting source material. Google confirms that its AI features retrieve relevant pages from its Search index through core Search ranking systems. Its guidance gives no organic position threshold, and it does not confirm the patent's positional measure for selection.

What to do:
1. List the questions prospects ask about the category, from sales calls, People Also Ask, and AI Mode follow-ups.
2. Check in Search Console which company page ranks for each one.
3. Improve the page that serves each question, and create one where a distinct need has no answer.

• ❌ Vendor A's per-entity pricing sits only in a gated PDF.
• ✔ Vendor A publishes a crawlable pricing page that ranks for "Vendor A pricing per entity" and states the price in text.

2. Selection rate and popularity: how often people choose a page. The patent names "a selection rate" of a document for the query, a selection rate across many queries, and "an overall popularity measure". Selection rate measures how often searchers choose a result. Google says Search uses "aggregated and anonymized interaction data to assess whether search results are relevant to queries".

The court record adds detail for Search. It found that Navboost, a "memorization system" that aggregates click-and-query data about the results Google shows, runs on 13 months of that data. Glue is the wider log of each query and how people interacted with the page. Pandu Nayak, a Google search executive, testified that "Glue is just another name for [N]avboost that includes all of the other features on the page".

On popularity, the court noted two exhibits suggesting it rests on "Chrome visit data" and "the number of anchors", the links between pages.

That is evidence about Search ranking, and the step from there to AI link selection is our inference.

What to do: Write titles and descriptions that state the specific answer the page gives, so the result earns the click for its question.

• ❌ Title tag: "Pricing | Vendor A"
• ✔ Title tag: "Vendor A pricing: per-entity plans billed on one invoice"

3. Trust: the author, the domain, and inbound links. The patent's trustworthiness measure is one "generated based on an author thereof, a domain thereof, and/or inbound link(s) thereto". Inbound links are what SEO calls backlinks. Google documents link analysis and PageRank as part of its core ranking.

A Google document quoted in the court record calls PageRank "a single signal relating to distance from a known good source", and it feeds Google's Quality score, a page-quality signal in Search ranking. Google's guidance asks whether it is clear who wrote a page, and it says trust matters most among the qualities it calls E-E-A-T (experience, expertise, authoritativeness, and trustworthiness).

A named owner with a profile helps readers judge who stands behind product, security, and compliance facts.

What to do:
1. Sign product, security, and compliance pages with a named owner and a profile page.
2. Earn links from relevant independent sources, such as partner directories, auditors, and industry publications.

• ❌ An unsigned security page with no named owner.
• ✔ The same page signed by Vendor A's head of security, with a profile page and a link from the auditor's directory.

4. Freshness: when a page was created or updated. Freshness, in the patent, "reflects recency of creation or updating". Google runs various "query deserves freshness" systems that show fresher content where newer material is expected. Its guidance also warns against changing a page's date to make it look fresh when the content has not substantially changed.

What to do: Update pages whose facts change, show the real date, and keep a short changelog on pricing, limits, and compliance pages.

• ❌ "Updated 1 October 2026" on copy that has not changed since 2024.
• ✔ A changelog line: "Updated 14 October 2026: entity limit on the Growth plan raised from 10 to 25."

5. Location and language of the person searching. The patent compares where a query comes from with the location a document corresponds to, and the query's language with the document's language. Its own example is a search for history of Louisville, which can draw on different source pages from Louisville, Kentucky, than from Louisville, Colorado. Google says Search uses location, past searches, and settings to decide what is most relevant.

Where a fintech company's licenses, offers, or prices differ by country, market pages can state those differences.

What to do: Publish market and language versions with hreflang where offers, licenses, or prices differ.

• ❌ One global page saying Vendor A is "regulated in several markets".
• ✔ A UK page naming Vendor A's authorization by the Financial Conduct Authority and a German page naming its BaFin license, linked with hreflang.

6. The person: profile, recent searches, and chosen sources. User-dependent measures rest on "attributes of a user profile", recent queries, and recent non-query interactions. In the stateful chat application, session context also shapes the synthetic queries written to retrieve documents.

Connected personal context can change what AI Mode recommends to a signed-in person who opts in, and a signed-out tracker does not reproduce it. Google let subscribers connect Gmail and Google Photos to AI Mode in January 2026 and extended these opt-in connections to nearly 200 countries and territories in May 2026.

One person-level lever is open to companies. Since May 27, 2026, Preferred Sources highlight websites a person has chosen inside AI Overviews and AI Mode, and "any website that publishes fresh content is eligible". Google says people are twice as likely to click a Preferred Source.

For AI answers, Google describes labels and highlighting, not a change in which sources are picked.

What to do: Answer the follow-up questions where they belong, and ask regular readers to add the website as a Preferred Source if the company publishes often.

• ❌ A product page that ends at the feature list.
• ✔ The same page linking the migration, security, and per-entity pricing answers a prospect asks next.

7. Information gain: what a page adds beyond what was read. Google's information gain patent, filed in 2018 and granted in 2022, scores each new document by "a quantity of new information" beyond what was already presented from earlier documents. In its first claim, an automated assistant presented that information, and a continuation was granted in 2025.

The description gives an example scale from 0.00 to 1.00 and says search-engine scores "may be adjusted" with it. Its training labels can come from people asked, "Was this document redundant?"

No Google source confirms an information gain score in Search or its AI features, and the claimed score is per person. Google separately documents original content systems and a diversity system that "generally won't show more than two web page listings" from one website among top results.

Google's guide also says a unique point of view helps, because its AI features look at a variety of sources. Google recommends original, useful information, while its information gain patent compares documents against one person's earlier information. Neither source says that repeating existing information blocks a page from being retrieved, cited, or shown in Search.

What to do: Put first-party numbers, worked examples, exact limits, and short case studies on the page, and cut paragraphs that restate the top results.

• ❌ "Five benefits of expense automation", restating the top results.
• ✔ Median approval time by plan from Vendor A's own product data, with the sample and dates.

Which AI Overview factors has Google confirmed, and which are open?

Set side by side, the patent measures and Google's statements leave a short list of confirmed steps for AI answers and a longer list confirmed only for Search.

Three open questions would change the advice. The first is what AI Overviews retrieves from. In the remedies opinion, the court found that Google grounds its Gemini models with FastSearch, which "generates abbreviated, ranked web results", retrieves fewer documents than Search, and is not used for Google's results page. It did not say whether AI Overviews uses FastSearch or Search's full ranking.

The second is how far AI Overviews and AI Mode differ, since Google says they "may use different models and techniques". The third is how many documents a live AI Overview uses. Google's product guidance gives no fixed count, while the patent's examples use set quantities such as five documents.

What happens when a CFO asks AI Mode to compare expense software?

The confirmed steps and the patent measures come together in one buying question. In this worked example with invented vendors, the chief financial officer (CFO) of a 200-person company in London asks AI Mode for expense management software. It has to work with NetSuite across several legal entities. The walkthrough combines separate patent proposals and Google's product descriptions into one hypothetical example.

The cited sources give no weighting formula for AI Overview citations, so the trace shows where a page can drop out, never a score.

In this hypothetical example, Vendor B has no accessible answer page for the searches the comparison needs, and Vendor C changes its displayed date without updating the content, a practice Google's guidance warns against. These details identify website improvements; they do not show which vendor a live AI answer will cite.

Vendor A controls its page content, authorship, and updates, while Google decides retrieval, ranking, and citation. Its review-platform listing and marketplace page also rank for two facets, so they can enter the pool as well.

The follow-up turn then changes the pool. It adds a question about approval policies for each legal entity, which extends the context available for retrieval, and in the stateful chat application, synthetic queries draw on the session's earlier queries. A page that answers approval policies per entity directly can enter the pool for the next citation.

A separate signed-out check in Google US on October 27, 2026, used the query expense management software that integrates with NetSuite for multiple entities. The answer excerpts we kept name six vendors, and the answer closes with two clarifying questions about the company's setup and needs. A clarification query is one of the variant types the query variants patent lists.

How to optimize a website for AI Overviews source selection

Start with the changes Google documents: eligible pages, clear product information, useful content, and accurate third-party representation. Google's guide also sets one limit. On separate content for every fan-out query, it warns that "doing so primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy".

Its systems also grasp relevance "even when there is no exact match" between the query and the page, so it makes sense to group variations that share a search intent on one useful page and to create separate pages only for distinct needs.

Nine steps cover the work, each with a test that says when it is done:

1. Check eligibility. In a technical SEO check, the key pages are indexed, allow snippets with no page-level nosnippet, and are included in generative AI features in Search Console, while passages meant to be quoted carry no data-nosnippet.
2. Map the questions. Each search intent behind a prospect question maps to the company page that best answers it, with the genuine gaps listed.
3. Answer each question on the page that serves it. The answer sits in text on that page, linked from the product pages, with no near-duplicate pages per phrasing.
4. Earn the click. The title and description state the answer the page gives.
5. Make trust visible. Product, security, and compliance pages carry a named owner with a profile page.
6. Update for real. Pages whose facts change show a real date and a changelog line.
7. Publish market versions. Each market with different offers, licenses, or prices has its own page with hreflang.
8. Add what the top pages lack. Each key page carries a first-party number, an example, or an exact limit the top results do not.
9. Be accurately present on third-party pages. The review, comparison, and marketplace pages that rank for facet questions describe the company correctly, with no inauthentic mentions.

The suggested priorities below focus on the product, commercial, and regulatory facts prospects need in each business model:

Read more: Ranking in AI Overviews

How to measure source selection. Four measures track eligibility, observed rankings, citations, and visibility in AI features:

A citation records a supporting link, and the pages the model read can differ, since the summaries patent describes linked pages that sit outside the results the summary was written from.

An AI visibility tracker can log AI Overview citations weekly, recording the product, prompt, locale, account state, and date of each check. Several runs are read before any page is rewritten, and finding the cause of a change needs further evidence.

How we read Google's filings at Omnius

Omnius reads a patent as the direction Google protected and acts first on the steps Google confirms. AI search changes "by model, context, wording, conversation history, location, retrieval, reranking, entities, and other variables", as the Omnius technology page puts it. So Omnius, as an SEO and AI search agency, works to raise the probability of favorable positioning, never to chase one answer. People decide what each page says, while data processing measures rankings and citations. The work runs in four parts:

• Facet map: the questions prospects ask, the searches they fan out into, and the company page that ranks for each.
• Page answers: one page per facet, with real updates, named authors, and market versions where offers differ.
• Eligibility: index and snippet checks and the generative AI setting.
• Dated checks: cited pages and organic positions, read over several runs.

The AI search audit assesses visibility, cited sources, brand representation, and technical access across AI search features.

FAQ

Does a page need to rank to be cited in AI Overviews?

Google retrieves supporting pages through Google Search, including related searches it writes for a question, and its guidance gives no required organic position. Google says AI Overviews are "built to only show information that is backed up by top web results". For example, a pricing page that ranks for a fan-out search on per-entity pricing can be cited for a broader software question.

What sources does AI Overview use, and where does it get its information?

AI Overviews use pages from Google's Search index, retrieved through its core ranking systems. AI Mode also draws on "fresh, real-time sources like the Knowledge Graph".

What is information gain in SEO, and is it a ranking factor?

Information gain is not a confirmed ranking factor, since Google has not named it as one. In Google's information gain patent, granted in 2022, the score measures new information relative to what one person has already been shown.

Why do AI Overviews and AI Mode cite different sources for the same question?

The two features can cite different sources because Google says they "may use different models and techniques". Google also describes AI Mode adjusting its search plan based on what it finds during earlier searches.

Can a website opt out of AI Overviews and AI Mode?

Yes, a website can opt out through a Search Console setting, available to all websites since August 31, 2026, and Google says the setting "will not be used as a ranking signal" for other results.

Does structured data help a page get selected?

Structured data is not required for AI Overviews or AI Mode, and Google says no special schema markup is needed, though it still recommends structured data for rich results. None of the selection passages in the six patent families names structured data.

Methodology

This guide reads six Google patent families in their primary text and checks each step against Google's own statements.

• Patents. The claims, descriptions, and bibliographic data of US11769017B1 and its four family members, US11354342B2 and its 2025 continuation, US20240289407A1, US12639300B2, US11663201B2 and its 2026 continuation, and US12158907B1, read on patents.google.com on October 27, 2026.
• Family members. US11886828B1 covers what goes into the prompt, US11900068B1 linking verified portions, and US12118325B2 when to use the model at all; the pending US20250005303A1 covers which model to use.
• Google's statements. Search Central documentation, Google blog posts from May 2024 to October 2026, and the US v. Google remedies opinion of September 2, 2025, quoted as the court records them.
• AI Mode answers. Two answers read on October 27, 2026, in Google US, signed out, one run each. The patent question is saved in full; the NetSuite expense software query is saved as answer excerpts with the titles of the organic results, an approximate time, and the locale.
• Screenshots. The patent pages and Google's guide were captured on October 27, 2026.
• Charts. The matrix records each concept apart from the operation it serves. A claim cell means a patent claim names the concept, even for another operation such as writing queries, and a Google cell means a Google page, post, or court exhibit says classic Search or an AI feature uses it. A blank cell means no such statement was found, and the timeline uses the dates Google Patents lists.
• Scope. The review covers the six named patent families and leaves Google's other retrieval and reranking filings out of scope.
• Limits. A patent shows a method Google protected, not what a product runs, and no source here tests whether a measure changes which pages AI Overviews cite. The worked example uses invented vendors.
• Publishers' interests. Google describes its own products, and the court exhibits quoted here are Google documents. Secondary analyses of these patents were read for coverage and are not cited.

Sources. The sources below are grouped by publisher.

• Google patents: US11769017B1, US11886828B1, US11900068B1, US12118325B2, US20250005303A1, US11354342B2, US12326889B2, US20240289407A1, US12639300B2, US20250117381A1, US11663201B2, US12579133B2, US12158907B1
• Google Search Central: AI features and a website, Optimizing for generative AI features, Ranking systems guide, Helpful, reliable, people-first content, Generative AI performance reports, A new resource for optimizing for generative AI
• Google blog and Search pages: AI Overviews: About last week, Generative AI in Search, Expanding AI Overviews and introducing AI Mode, AI Mode update, May 2025, Personal Intelligence in AI Mode, Search at I/O 2026, 5 new ways to explore the web, Preferred Sources in AI Search, New controls for website owners, AI Mode and AI Overviews updates, AI-organized results, Web Guide, How Search works: ranking results
• Court record: United States v. Google LLC, remedies memorandum opinion, September 2, 2025