Quick answer: The Google AI Overview patent (US11769017B1) describes how an AI search summary picks a small set of source pages, checks each claim against them, scores its own confidence, and adapts to the person asking. It is a filing, not a confirmed spec of AI Overviews, but it is the most detailed public look at the mechanics.
The Google AI Overview patent is a rare primary source in a field full of guesswork. Most advice about AI Overviews is reverse engineered from outputs: someone screenshots a result, spots a pattern, and publishes a theory.
This article starts from the other end, with a document Google actually filed. It will not hand you a ranking formula. It does show where the real levers sit and where the limits are. We cover six mechanics from the filing, and we stay on how the system is described to work. For the tactical side, see our guide to getting cited in AI Overviews in Thailand.
What Is the Google AI Overview Patent?
US11769017B1, “Generative summaries for search results”, was filed by Google on March 20, 2023 and granted on September 26, 2023. It describes using a large language model to write a natural language summary for a query, feeding the model content from search result documents instead of the query alone.
The stated goal is practical. Fewer inaccurate answers, and fewer answers that are too detailed or too vague for the person asking.
Read it with two cautions. A patent describes what a system can do, and this one is full of optional phrasing. And Google has not said AI Overviews works exactly this way. Treat it as the best public blueprint, not a confirmed specification.
The takeaway: use the Google AI Overview patent to understand design logic and direction. Do not chase any exact threshold in it.
Where Do Source Pages Come From in the Google AI Overview Patent?
Not only from the query you typed. The patent describes assembling a set of source pages from up to four places:
- The query itself. Often just the top two or three responsive pages.
- Related queries. Queries people tend to issue close together in time.
- Recent queries. Your own recent searches, when they overlap topically.
- Implied queries. Queries Google generates from context or profile data.
The extra sources are not always pulled in. The filing says related, recent and implied sources are added selectively, for example when the pages for the main query are low quality or too similar to each other, and when the link between the queries is strong enough. The model then reads passages from the chosen pages and writes the summary.
The takeaway: a page can feed a summary for a query it never targeted. A reasonable reading is that depth around a topic gives you more ways in than a single page does, though the patent does not promise that.
How Does Google Pick Which Pages Get Used?
Candidates are filtered by three families of signals, and none of them is exotic.
- Query-dependent: position in the results for that query, how often the page gets selected for it, and how well its location and language match the query.
- Query-independent: trustworthiness (author, domain, inbound links), overall popularity, and freshness.
- User-dependent: how the page relates to the person’s profile, recent queries and recent activity.
Notice what is missing. There is no special “AI format” signal. Selection leans on the same trust, freshness and relevance signals as ordinary ranking, and position is one input among several, which is why a page ranking fourth can still be used. Language and location are first-class signals, and that matters in Thailand, where Thai and English queries behave like separate markets.
The takeaway: the fundamentals still gate everything. Our guide to Google ranking factors in 2026 covers the ones that move the needle.
How Does the Google AI Overview Patent Verify a Summary?
With a verification step, which is the least discussed part of the filing.
Each source is tagged with an identifier in the text fed to the model, so the output can tie statements back to the page behind them. Then comes a second check. The patent describes comparing a mathematical representation (an embedding) of a summary sentence against one for a passage in the candidate page. If the two are close enough, the page counts as verifying that sentence and gets linked to it.
Every portion also gets a confidence measure. It draws on the model’s own confidence, how many sources verify the claim, and how trustworthy those sources are. Four trusted pages agreeing beats one.
One more detail matters. The summary can include statements that come from the model’s training and cannot be derived from the sources. The retrieved content shapes them, but does not guarantee them. So a linked sentence is supported by a page, and an unlinked one may be the model talking.
The takeaway: a claim that several trusted pages agree on is easier for this check to verify than one only your page makes. Consistent facts across the web are not a branding nicety here. They are what the check looks for.
Can Low Confidence Switch the AI Summary Off?
According to the patent, yes. Confidence decides how the result is shown:
- High confidence: the summary can show first, with classic results held back or hidden behind a control.
- Medium confidence: the summary shows alongside regular results.
- Low confidence: the summary is suppressed and only regular results appear.
The filing goes further. It describes choosing among generative models per query, including a creative one, an image one, larger and smaller language models, or none at all when ordinary results already do the job.
The takeaway: a missing AI Overview can be a designed outcome, not a glitch. In our test of 100 Thai queries, Overviews appeared on 84, so they are common without being universal.
Why Do Two People See Different AI Overviews?
Because the patent says they can. Different submissions of the same query can produce different source sets, and so different summaries. The filing’s own example: a search for the history of Louisville from Louisville, Kentucky pulls different pages than the same search from Louisville, Colorado.
It also describes two adaptive behaviors:
- Familiarity. If a profile suggests you already know something, the summary can be written assuming you do.
- Revision after a click. After you open a result, the summary can be regenerated to skip what you just read and go deeper on what is left.
The takeaway: there is no single AI Overview ranking to check. Visibility is a distribution across locations, languages and users, so measure it as a sample. Our guide to measuring AI search traffic shows what GA4 and Search Console can and cannot tell you.
What Does the Google AI Overview Patent Mean for Your Content?
- Earn the ranking first. Trust, popularity and freshness gate source selection, so weak pages rarely get the chance to be quoted.
- Cover the neighborhood of a topic. Related-query pulls favor sites with real depth, not a lone page.
- Keep your facts identical everywhere. Verification rewards claims that independent pages corroborate.
- Test in both languages and more than one location. One screenshot proves nothing when summaries vary by person.
Our view at Inspira Digital Agency is simple. The Google AI Overview patent is a reason to stop treating AI Overviews as a separate game. Almost every selection signal in it is a classic SEO signal. The new pieces are verification, confidence gating and personalization, and all three reward being right and being corroborated over being clever. For the full framework, read our guide to AI search optimization for Thailand businesses.
Frequently Asked Questions
What is Google patent US11769017B1?
It is the Google AI Overview patent, titled Generative summaries for search results, filed in March 2023 and granted in September 2023. It describes using a large language model, fed with content from search result pages, to write a natural language summary for a query, with source verification, confidence scoring and personalization.
Is the patent proof of how AI Overviews works?
No. A patent describes what a system can do, often with optional wording, and Google has not confirmed that AI Overviews follows it exactly. Treat it as a credible blueprint for design logic, such as selective source gathering and confidence gating, not as a ranking formula.
Do AI Overviews only use the top-ranking pages?
Not necessarily. The patent describes selecting a small set, often the top two or three results for the query, and optionally adding pages from related, recent or implied queries. Position is one signal among several, alongside trust, popularity, freshness, location and language.
Can I optimize for a specific patent?
Not directly. The useful move is to act on the principles the Google AI Overview patent shows: build trust and freshness, publish facts that other sources corroborate, cover related questions, and measure visibility as a sample across languages and locations rather than as one fixed ranking.
The Bottom Line
If you take one thing from the Google AI Overview patent, take the shape of the system: gather sources, verify claims, score confidence, then show or hide. Each step rewards pages that are trusted, fresh and corroborated, and punishes pages that only sound confident. That is good news, because it is the same work that wins in classic search. You now know why a summary can cite a page you did not expect, and why no two people see exactly the same one. The next step is checking where your own pages stand.
Related reading: Traditional SEO vs. GEO: What is the Difference? and Get Cited by ChatGPT in Thailand: What Actually Works
Want to know whether Google’s AI answers cite your business or your competitors? Inspira Digital Agency runs AI search optimization in Bangkok, starting with a free AI visibility audit.
