How Google AI Overviews choose their sources
An AI Overview is a custom Gemini model grounded in Google Search: it retrieves pages from Google's own index, writes a summary above the results, and cites the sources it leaned on. Index inclusion gets you considered, query fan-out widens the field, and extraction plus trust decide whose link appears.
Key takeaway: AI Overviews cannot cite what Googlebot has not indexed; there is no separate AI crawl. From the indexed pool, Google decomposes the query into parallel sub-queries (fan-out), then the model cites the pages it can lift a clean passage from and trusts enough to stand behind. Ranking helps, but a well-structured, trusted answer often beats a higher-ranked wall of prose.
What actually produces an AI Overview?
An AI Overview is not a chatbot answering from memory. It is a custom Gemini model grounded in Google Search: for queries where Google decides an AI answer helps, the model retrieves live pages from Google’s own index, synthesises them into a summary at the very top of the results page, and attaches a cluster of citation links to the sources it leaned on.
That grounding is the whole story. Unlike engines that lean on a chatbot’s frozen training or a rival search index, AI Overviews run on the same Google index you have always optimised for. The work you have banked in Google carries over. The catch is that ranking and getting cited are no longer the same thing.
Mechanic 1: index inclusion is the entry ticket
AI Overviews can only cite what Googlebot has already crawled and indexed. There is no separate AI crawler to allow or block. If you are not in Google’s index for the topic, you cannot be a source, full stop.
This has a counterintuitive corollary about robots controls. Google-Extended, the token many sites blocked to opt out of AI training, is not what governs Overview citations: blocking it does not, by itself, pull your page out of AI Overviews, because those are sourced from the standard Search index. The control that genuinely keeps you out is the snippet family: a page carrying nosnippet or max-snippet:0 gives Google nothing it is permitted to lift into the summary. The trade-off is blunt, because the same rule strips your featured snippets and shortens your normal result snippet.
Mechanic 2: query fan-out widens the field
Google quietly decomposes one question into several related sub-queries, runs them in parallel, and stitches the answer from whichever pages best satisfy each strand.
This changes who can win. A page can be cited for a narrow sub-question it answers cleanly, “what does X cost?”, “is X safe for Y?”, even when it does not rank first for the headline term. Conversely, the page that tops the headline query can be absent from the Overview because it never directly answers any individual strand. Fan-out is why topic coverage, answering the cluster of questions around your subject, beats optimising one page for one phrase.
Mechanic 3: extraction decides who gets quoted
From the candidate pages, the model favours the ones it can lift a clean, direct passage from. A section that opens with a self-contained answer under a question-shaped heading is extraction-friendly; a correct answer spread across five paragraphs of build-up is not, and usually loses to a source the model can quote without rewriting.
The scale of this gap is measurable. Across 850,000+ sites in SearchScore’s SAVI benchmark, technical foundations average 70.1/100 while on-page structure averages 23.1/100. The sites are built and indexed fine; they are simply not quotable. That mismatch, rankable but not liftable, is the single most common reason healthy sites never appear in Overviews.
Mechanic 4: trust decides who gets named
Extraction narrows the field; trust picks the winner. The model favours sources it deems authoritative enough to stand behind: named authors with credentials, Organisation and author schema that resolve the entity cleanly, reviews and third-party references, and topical authority in Google’s existing sense. On YMYL topics (health, finance, legal) this bar is sharply higher, and thin or anonymous content is effectively excluded regardless of structure.
Entity resolution matters more than most sites realise: if structured data does not make clear who you are and what each page answers, the citation can go to a similarly named brand, or to nobody.
What this means for how you optimise
Two independent things now have to be true. You have to be rankable: indexed, relevant, trusted, the classic job. And you have to be quotable: structured so a single passage answers the question outright, with the trust signals to back it. Most sites have only done the first job, which is why a healthy rankings dashboard coexists with total absence from the answer users actually read.
The practical checklist:
- Key pages indexed and server-rendered, no JavaScript-trapped content
- No
nosnippet,data-nosnippetormax-snippet:0on pages you want cited - Every important section opens with a direct, liftable answer under a question-format heading
- The sub-questions around your topic each answered explicitly (fan-out coverage)
- Organisation, Person and author schema resolving your entity cleanly
- Real E-E-A-T: named authors, credentials, reviews, third-party references
The free Google AI Overviews Visibility Checker inspects all six on any URL, index inclusion, snippet eligibility, citable structure, fan-out coverage, authority and entity clarity, and returns a ranked fix list in about 60 seconds.
How do AI Overviews differ from the other engines’ citation systems?
- Gemini (in its own app) grounds on the same Google index, so the two rise and fall together; see how Gemini decides which websites to cite.
- ChatGPT blends frozen training data with live browsing that leans on Bing.
- Perplexity runs its own crawler and re-retrieves at every query, with freshness weighted hard.
- Claude answers from training by default, with an optional live web search.
AI Overviews are unique in one respect: they are the only AI answer surface whose entry ticket is literally your existing Google indexation, which makes them the most direct extension of classic SEO, and the clearest demonstration that classic SEO alone is no longer enough.
Related articles
- Google AI Overviews: the complete guide →
- How to appear in Google AI Overviews →
- Why your site isn’t in AI Overviews →
- What is GEO? The complete guide →
Sources & Further Reading
- SearchScore SAVI Report, April 2026 (850,000+ sites audited)
- Google Search Central – AI features and your website
- Google Search Central – Creating helpful, reliable, people-first content
- Academic research – GEO: Generative Engine Optimization (Aggarwal et al., arXiv)
Frequently asked questions
Does blocking Google-Extended remove my site from AI Overviews?
No. Google-Extended is a training and grounding opt-out for Google's generative models; AI Overview citations are sourced from the standard Search index built by Googlebot. The control that genuinely keeps a page out of the summary is `nosnippet` or `max-snippet:0`, at the cost of your featured and normal snippets too.
Can a page be cited in an AI Overview without ranking first?
Yes, routinely. Query fan-out means Google runs several related sub-queries in parallel and cites whichever pages best satisfy each strand. A page that answers one strand cleanly can be cited from deep in the rankings, while the headline number-one goes unnamed.
Why does the Overview cite my competitor when I outrank them?
Because citation is decided by extraction and trust, not position. If their page states the answer in one liftable sentence with a named, credentialled author while yours is a well-ranked wall of prose, the model quotes them. The fix list, in priority order, is in why your site isn't in AI Overviews.
Do AI Overviews appear for every query?
No. Google shows an Overview where it judges a generated answer helps, most often for informational and comparative questions, and presence for a given query can change over time. That is why sensible measurement uses a basket of queries tracked over weeks rather than a single search: you are optimising your eligibility and citability, not toggling a switch.
Does adding schema guarantee a citation?
No single signal does. Schema resolves your entity and your page's purpose so extraction and attribution can work, but the citation is still decided by passage quality and trust. Think of structured data as removing ambiguity that would otherwise disqualify you, not as a ranking lever in itself.
Are AI Overviews the same as featured snippets?
No, though they occupy similar territory. A featured snippet lifts one passage from one page; an AI Overview is generated by a Gemini model synthesising several sources, with a cluster of citations. The overlap is the mechanics: both depend on Google being permitted to quote you (the snippet controls) and on your page containing a passage that answers the question outright, which is why the same structural work feeds both.