How Do AI Overviews Choose Sources?

From Google's own search index — but not the way most people picture it. An AI Overview is not a summary of the ten blue links below it. Google documents that its AI features use "query fan-out" — issuing multiple related searches across subtopics and data sources — and that during generation its models "identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links associated with the response than with a classic web search."

Decode those two sentences and you have the whole sourcing model. Your one query becomes many background searches. Each background search retrieves its own candidate pages. The Gemini model drafts an answer constrained to what those retrieved passages actually say — a process called grounding — and the pages whose passages back the final claims become the citations. So the sources are always drawn from Google's indexed, retrievable web (nothing outside the index can be cited), but from a far wider pool than the visible page one. The full pipeline, stage by stage, is mapped in how AI Overviews work.

Increasingly, no — and this is the most misunderstood fact in AI search. Ahrefs analyzed 863,000 keyword SERPs and 4 million AI Overview URLs (March 2026) and found the citation pool splits almost evenly into thirds:

Eight months earlier, roughly 76% of citations came from the top 10. That overlap has collapsed to about 38% — meaning nearly two-thirds of the sources Google's AI now cites are not on page one for the query you typed. They're pages that won one of the invisible fan-out searches instead.

Two business implications follow directly. Ranking #1 no longer means you own the citation — we routinely audit clients who rank first and are absent from the Overview above their own listing. And ranking on page four no longer means you're out of the game — a page that cleanly answers one specific sub-question can be cited above sites that outrank it everywhere. Selection has decoupled from position.

Mostly — with documented, instructive failures. On the credible side: every cited source has passed through Google's ordinary ranking and quality systems just to be retrievable, and answers are typically triangulated rather than single-sourced — Pew found 88% of AI summaries cited three or more sources. The grounding step exists precisely to tie the model's claims to retrievable pages instead of its imagination.

But the system judges credibility statistically, not editorially — and that gap has produced famous misses. In May 2024, AI Overviews told users to put glue on pizza and eat a rock a day, grounding those answers in a decade-old Reddit joke and a satirical article; Google restricted the feature and filtered humor content in response. The lesson cuts both ways. For users: treat Overviews as a well-sourced starting point, and check the citations on anything that matters. For businesses: the bar you must clear is machine-legible credibility — verifiable claims, cited evidence, corroborated identity — because that is what the selector can actually measure.

— "The Overview doesn't cite the best page; it cites the page it could verify fastest. Retrieval finds you, grounding tests whether your sentences hold weight, and corroboration decides if you're safe to name. Win those three machine judgments and your rank barely matters — lose them and #1 won't save you." — Thomas, Founder of AISEO USA, 16 years in digital marketing

Stop auditing rankings and start auditing citations — they are now different scoreboards with only ~38% overlap. Pull the AI Overviews for your money queries and list who is actually cited; that list is your real competitive set, and it will include domains your rank tracker has never shown you. Then work the selection criteria in order: retrievable, quotable, corroborated, current. The corroboration layer — third-party mentions and entity trust — is the slowest to build and the most durable, which is why it anchors our generative engine optimization services and the broader AI SEO program. To see today's picture in one pass, our free AI visibility audit maps which sources the Overviews on your queries currently trust — and whether any of them are you.