How AI Overviews Choose Sources: The Citation Layer in Google’s Patents

Last updated 30 October 2026. Patent statuses checked October 2026.

Google's AI Overviews write the answer first, then check it against documents. Documents that agree with an AI Overview sentence get that sentence's link. That verification step is the citation layer. Most AI search advice never mentions it.

Definitions

The citation layer is the step where an AI Overview links each sentence to documents that confirm it. US11769017, "Generative summaries for search results," is Google's AI Overviews patent. Google filed it in March 2023 and it was granted in September 2023.

Key takeaways

In AI Overviews, verification is a second retrieval keyed to the answer's sentence, not the user's query.

Google's US11900068 claims one AI Overview sentence verified by two documents, each with its own link.

FAQ schema and question headings barely predicted AIO citation in Rankera's 30,613-overview study (AUC 0.51).

Track AI Overview citations per sentence. Shared links mark corroborated claims. Solo links mark claims only you make.

Related readings

Google revealed nothing

In October, a viral post claimed Google had "accidentally revealed" how its AI search picks sources. The X version passed 53,900 views. It summarised Metehan Yesilyurt's teardown of Vertex AI Search, Google Cloud's enterprise retrieval product.

Seven signals, a 500-token chunk cap, four pipeline stages. The advice that followed was familiar. Shorter sections, question headings, cleaner HTML.

Everything in that teardown describes extraction. Extraction is the earlier of the two steps this essay covers.

The top reply on the LinkedIn version came from Mark Williams-Cook. He pointed to US11769017, Google's patent describing how AI Overviews choose documents and citations. Google filed it in March 2023. It has been public since September 2023.

That exchange is the whole problem with how the industry reads Google. We reason from the newest artefact, not the one closest to the system. A Cloud product sold to enterprises becomes a window into Search. A leaked attribute list becomes a weighting table.

Until this week, most coverage stopped at the abstract. On 28 October, Omnius walked the patent family's selection steps. Nobody has yet ranked the evidence behind these claims.

Read past the abstract and the advice changes. Formatting helps a page get read. Agreement gets it credited.

Six kinds of evidence, six different claims

Every "Google revealed" post rests on one of six evidence types. Each proves something different. Most posts borrow the authority of the strongest type for a claim built on the weakest.

Evidence type	Example	Proves	Does not prove
Patent claims	US11900068, claims 1, 12	What Google chose to protect	That it ships today
Patent description	Source selection measures in US11769017	A design Google disclosed	Production use or weighting
Google statements and docs	Search Central, Gary Illyes talks	Stated production behaviour	Internal weights
Content Warehouse leak	`siteRadius`, `queriesForWhichOfficial`	Fields Google stores	That ranking uses them
Google Cloud product docs	Vertex AI Search, Discovery Engine	How a sold retrieval product works	That Search shares the stack
Third-party measurement	Graphite, Resoneo, Ranko audits	Behaviour in a sample	Mechanism or cause

Two rules follow. A patent claim outranks a patent description, because claims are what Google chose to defend. A Cloud product doc proves how that product works, and nothing more.

Google's own statements confirm less than most posts imply. Its AI features documentation says AI Overviews may use query fan-out. It also requires a page to be indexed and snippet-eligible to appear as a supporting link. It says nothing about how Google chooses among eligible pages.

The Vertex teardowns show why the second rule matters. Discovery Engine lets each customer set its own boost and bury rules. Reading those as Google Search logic misreads who holds the controls. One agency write-up still presented the signals as documented methodology rather than inference.

Patents carry a known limit too. Google patents systems it never ships, and AI Overviews have changed since 2023. A patent shows a design worth protecting. It remains closer to Search than a product built for other companies' sites.

How do AI Overviews pick citations?

US11769017 splits the AI Overviews pipeline in two. In AI Overviews, generation comes first and citation comes after.

Generation. The system selects a small set of search result documents. The patent's example is two or three. Their content goes to the LLM, which writes the summary. The patent lets that summary include material from the model's own training.

Those few documents are chosen by three kinds of measure in the patent's description. Query measures include ranking position, click selection rate, locality and language. Query-independent measures include selection rate across queries, trust from author, domain and inbound links, popularity and freshness. User measures draw on profile and recent activity.

When those core results are weak or too similar, the patent pulls in documents from related queries. In US11769017, related queries are ones the same device or account issues close in time. That gate reads as an early form of what Google now calls query fan-out.

Verification. The system then splits the summary into portions, often a sentence each. For each portion it looks for a document that verifies it. The described method compares an embedding of the portion with embeddings of document passages. Inside a distance threshold, the document verifies the portion and the portion gets a link.

The link goes to whichever documents confirm the sentence

Generation: query arrives → top 2 to 3 search results selected → LLM writes the summary

Verification: summary split into portions → each compared with passage meaning (a fresh search on the portion can reach docs the LLM never read) → every confirming doc linked → confidence sets display (high: hide results, low: drop summary)

The citation link is decided at verification, not generation. A sentence can carry several links when several documents confirm it.

Dan Petrovic at DEJAN mapped this linking engine earlier this year. Two details in it matter for anyone chasing citations.

First, the verifying document need not be one the model read. The patent describes running a fresh search on the summary portion itself, then checking those results. An AI Overview can credit a page for a sentence another source supplied.

Second, the patent's own interface example shows cited sources that differ from the top three organic results. Citation and ranking are separate selections.

Three independent datasets put numbers on that gap. Their methods differ, so read them as a range.

Study	Sample	Finding
Conductor, Sep 2026	167.9M US AIO citations	57% of citations sit outside the organic top 10
WashU, Sep 2026	55,393 searches	29.8% of cited domains never appeared on page one
Rankera, Oct 2026	30,613 AIOs, 10 countries	72% of #1 pages cited, yet 92% of AIOs cite beyond the top 20

So the citation question is narrower than most advice assumes. Does the page hold a passage that independently confirms what the answer already says?

Corroboration is in the claims

The strongest support for this argument sits in a continuation even recent coverage reads narrowly. US11900068, granted February 2024, claims the multi-source case directly. Omnius describes it as covering linked, verified portions of an answer. The claim language says more.

US11900068's first independent claim covers one AI Overview sentence verified by two different documents. Each verifying document gets its own selectable link, placed immediately after that sentence. A sentence carries several sources because several sources confirm it.

That is claimed language. In US11900068, Google protected verifying one AI Overview sentence against several documents.

The parent patent adds a confidence layer. Its claim 13 covers a confidence measure for each portion, shown as an annotation. The description says how that measure moves. Confidence rises when more documents verify a portion, and again when those documents are trustworthy. In Google's AIO patent, four trusted verifying sources score higher confidence than one.

The description then ties confidence to display. High confidence can show the summary with organic results held back. Low confidence can suppress the summary and show only results.

Read together, corroboration does two jobs. It earns the link on a sentence. It also helps decide whether the summary appears at all.

Independent measurement points the same way. Ranko audited 5,640 multi-citation AI Overview cards. Where cited sources disagreed on a fact, the majority value won about 68% of resolved conflicts.

In our Proof of Importance model, corroboration is signal seven. The patent family gives it a mechanism.

Does chunking help AI Overview citations?

Chunking helps AI Overview citations only at the extraction step. It is hygiene, and it works on the earlier step.

The parent patent describes picking passages by query term overlap or embedding distance to the query. Clear sections and question headings help that step. They make the right passage easy to lift.

Position matters at this step too. Rankera found the quoted passage sat in a page's first quarter 52% of the time, against 32% by chance. AirOps reports 55% of AIO citations come from a page's top 30%.

The Vertex teardown's 500-token chunk cap belongs here too. It governs how a passage gets lifted, which the patent places before verification.

But extraction only gets a page considered. Verification asks a harder question. Does this passage confirm a claim the summary already makes?

A well-chunked page making a claim nobody else makes has little to confirm. A plain page stating the agreed fact in clear terms can still verify a sentence. Structure decides whether you get read. Agreement decides whether you get credited.

Rankera's October study puts numbers on this. FAQ schema, TL;DR blocks and question headings barely predicted citation, with an AUC of 0.51 to 0.52. An AUC of 0.50 is a coin flip. Meaning match scored 0.56 and recency 0.55. The study is one snapshot and shows correlation only.

Bridgehead's September review of about 350 sources lands nearby. It rated the evidence for rewriting pages to be "citable" as weak.

The retrieval objection

The same review argues retrieval is the real bottleneck. For most pages that is right. A page never retrieved cannot be cited.

The patent complicates the objection in one way. Verification can run its own search on the summary portion. That is a second retrieval, keyed to the answer's sentence rather than the user's query. A page stating the answer's facts in the answer's terms gets two chances to be found.

So the order for a SaaS team is retrievable first, corroborated second. Any decent template solves most of the first. The second lives mostly off your site. Aleyda Solís found 83.6% to 93.4% of citation weight sits off-domain across B2B SaaS subverticals.

What I'd do

For a funded SaaS team, the work shifts from page formatting to claim distribution.

Audit claims, not pages. List the ten facts you want AI Overviews to state about your category and product. For each, count the independent, trusted pages that state it the way you do.
Write claims in the answer's terms. One fact per sentence, each able to stand alone. Use the words a summary would use. Verification compares meaning, so plain phrasing sits closer than clever phrasing.
Seed the corroboration. Get each claim stated on review sites, comparison pages, partner docs and customer write-ups. Include video: YouTube took 14% of AIO citations in Rankera's study. Each mention must stand on its own page or video.
Sequence original data. Proprietary numbers earn information gain, but nobody can verify them on day one. Publish first, then get them repeated by credible third parties.
Track citations per sentence. Record which AI Overview sentence carries your link and which domains share it. Shared links mark corroborated claims. Solo links mark claims only you make.
Keep chunking in QA. Clean structure stays on the pre-publish checklist. It stops being the strategy.

FAQs

How do Google AI Overviews choose which sources to cite?

Google's patent US11769017 describes two steps. The system writes a summary from a few top results. It then links each sentence to documents that confirm it.