A Google patent has been doing the rounds with headlines about a leak, a slip, an accidental reveal. Begin with the correction, because it changes how you should read everything that follows: nothing here was accidental. Patents are published deliberately. Google’s lawyers drafted this one, paid the fees, and prosecuted it through to grant. Publication is the point of the system, not a mistake in it.
What is actually true is more interesting than the leak framing. In US 11,769,017 B1, “Generative summaries for search results”, Google set down in writing how it thought a search engine should assemble an AI summary. It did so with a priority date of 30 December 2022, before Bard was announced and well over a year before AI Overviews reached most people. That is a real window into what Google’s engineers considered worth owning, at a moment when almost nobody outside the company had seen what was coming.
It is also not a ranking factor list, and reading it as one is a mistake that ends up in client decks. This post covers what the document describes, what it cannot tell you, and how to read a search patent without talking yourself into something.
The verified details
| Detail | Information |
|---|---|
| Patent | US 11,769,017 B1, application 18/123,861 |
| Title | Generative summaries for search results |
| Assignee | Google LLC |
| Priority date | 30 December 2022 |
| Filed | 20 March 2023 |
| Granted | 26 September 2023 |
| Status | Active, anticipated expiry 20 March 2043 |
| Family | Includes PCT/US2023/086339 and EP23848676.5A |
The named inventors are Matthew K. Gray, John Blitzer, Corinn Herrick, Srinivasan Venkatachary, Jayant Madhavan, Sam Oates, Phiroze Parakh, Aditya Shah, Mahsan Rofouei and Ibrahim Badr.
What the document describes
The core idea is selectively using a large language model to generate a natural language summary shown in response to a query, where the model processes content beyond the query itself. The stated purpose is to reduce inaccuracies and the risk of a summary that is too general or too specific.
The mechanics described include the following.
- Inputs beyond the query. The query may be typed, spoken, image-based or implied. Alongside it, the system can feed in content from search result documents that are responsive to the query, plus related queries, recent queries or implied queries.
- How candidate documents are chosen. Selection is described using query-dependent, query-independent and user-dependent measures, with examples including ranking, selection rate, locality, language, trustworthiness, freshness and user profile attributes.
- Fitting the content in. Document content may itself be summarised first to fit within the model’s memory limits.
- Verification by linking. Portions of the summary can be linked to documents that verify them. Verification is described either from source identifiers in the model’s output, or by comparing embeddings and testing the distance against a threshold.
- Confidence measures. These are described as drawn from the model’s own confidence together with the number and trustworthiness of supporting documents, and may be surfaced as colour annotations or labels such as “high confidence”.
- Regeneration after interaction. If a user interacts with the search results, the summary can be regenerated to account for what they have already seen.
- Familiarity from the profile. The prompt can be adjusted when a profile suggests the user already knows the material.
- Model selection. The system may select none, one, or several generative models for a given query.
Four reasons this is not a ranking factor list
Filed is not shipped. The priority date is 30 December 2022. That is before Bard was announced and well over a year before AI Overviews rolled out broadly. Large companies patent options, including options they never build. A filing tells you what an engineering team considered worth protecting at a moment in time, which is a different question from what runs in production today.
The description is not the claims. A patent has two parts that matter very differently. The claims are the legally operative scope. The description is deliberately expansive, written in “may” and “can”, and drafted to cover as many variations as possible so that competitors cannot design around it. Almost every SEO article about a patent quotes the description.
Being straight about our own limits here: in preparing this post we could not read the granted claims in the public rendering of the patent, only the description. So everything in the section above is what the application describes, not what is legally claimed. We would rather say that than imply a precision we did not have.
“Can” is not “does”. The document describes confidence shown as colour annotations and labels. Nobody has seen that in AI Overviews. Described features are a menu, not a changelog.
One patent is not the system. This one sits in a family that includes an international application and a European filing, and Google holds a very large number of related filings. Reading one document and describing it as how AI search works is like reading one function and describing it as the application.
What is still worth taking from it
Three things, held loosely.
The vocabulary of document selection. Ranking, selection rate, locality, language, trustworthiness and freshness are the measures named for choosing which documents feed the summary. None of that is a revelation, and that is rather the point: it matches Google’s public documentation and what practitioners already observe. When a patent, the documentation and observed behaviour all point the same way, the direction is worth some confidence. One of them alone is not.
The structural implication is the useful part. In the described method, the summary is generated from documents that were already retrieved as responsive to the query. If that is how it works, then classic retrievability sits upstream of everything else. A page that cannot be found and is not eligible to be shown cannot be summarised. We set out the practical version of this in how to appear in Google AI Overviews.
Verification rewards plain statements. If a portion of a summary is verified by comparing embeddings against candidate documents, then a page that states a claim directly is easier to verify against than one that circles the subject for four paragraphs before arriving. That is the same conclusion we reach from watching what gets cited, which is why it is worth acting on even though the patent alone would not justify it.
How to read a search patent without fooling yourself
- Read the claims first. If you only have access to the description, say so when you write about it.
- Check the priority date against the product timeline. A 2022 filing describing generative summaries predates the products people are comparing it to.
- Check the family. Continuations and international filings often tell you more about what the company is still pursuing than the original grant does.
- Never present a patent as a confirmed ranking factor. Not in a deck, not in a proposal, not in a post. It is the fastest way to lose a technically literate client.
- Use it as a hypothesis generator. If the patent suggests freshness and trustworthiness matter for selection, go and test it against your own question set rather than quoting the filing as evidence.
The part that does not change
Here is the uncomfortable conclusion for anyone hoping a patent contains a shortcut. Whether or not this document describes what runs in production, the work is the same. Be retrievable. Say the thing plainly, early, in text a model can lift. Be corroborated by sources other than yourself. Keep the page accurate and current.
That list does not change if the patent is implemented exactly as written, and it does not change if it was never built. Which is a reasonable test for any piece of SEO analysis: if the recommendation would be identical either way, the evidence was decoration.
Source: US 11,769,017 B1, Generative summaries for search results, Google LLC, on Google Patents. Bibliographic details and the described method are taken from that page. Google Patents notes that priority date and expiry are estimates rather than legal conclusions.
