Skip to main content

How Google AI Overviews Choose Sources

By Edgar Li•

Google AI Overviews are grounded in Google's own search index, so a page generally has to be indexed and eligible to appear in Search before it can be cited. Beyond that, Google does not publish how sources are chosen. It says its AI features may issue several related searches for one query, which it calls query fan-out, and nothing in its tools shows you those searches. What outside studies add is a partial picture: ranking on the first page helps, but a large share of cited pages come from outside the top ten.

Key Takeaways

  • Google's documentation says AI features use its index and standard Search requirements, with no special markup needed.
  • Google runs several related searches per question. Any tool that shows you Google's fan-out queries is simulating them.
  • Studies of how much of an Overview comes from the top ten range from about 17 percent to about 52 percent, depending on the vendor, the year and the topic.
  • A sandbox site removed from Google's index vanished from its AI surfaces in one test.
  • Controls exist: Google offers snippet controls and a Search Console setting that removes a site from AI Overviews and AI Mode without removing it from Search.

What Google Says

Three points come from Google's own pages (AI features documentation, updated December 2025; AI optimisation guide, updated July 2026).

  1. Indexing and snippet eligibility are the entry ticket. To be included as a supporting link, a page must be indexed and eligible to be shown with a snippet. There is no separate application.
  2. Query fan-out. AI Overviews and AI Mode "may use a query fan-out technique", issuing multiple related searches across subtopics and sources to build the response.
  3. No special markup. Google says you do not need structured data, a new file or specific chunking to appear. See does schema markup help AI citations.

Google's page does not give a ranking formula, a list of factors or a weight for any signal. Anything claiming to reveal one is inference.

What Studies Find About Ranking

Several vendors measured how many AI Overview citations come from pages that also rank in the classic top ten for the query. The figures differ because the methods differ.

Vendor, dateWhat it countsResult
BrightEdge, Feb 2026Cited sources ranking top 10, Feb to Jul 2025About 17%
Conductor, Aug 2026AIO citations from top-10 pages, May to Jun 202642 to 52% across publishing, insurance and nonprofit
seoClarity, Jul 2025Citations from top-20 pages, 362,000 keywords56%

These cannot be combined into one number. They cover different periods, industries and definitions of "top". They do agree on direction: ranking is related to being cited, and a meaningful share of cited pages rank nowhere near the first page. Otterly (August 2026) reported that a sandbox site Google removed from its index disappeared from Google's AI surfaces, which fits Google's statement that the index is the foundation.

A further complication is that rankings and citations are measured at the page level against the query. The Overview may be built from a related search you never see. A page that does not rank for the headline question can still be pulled in for a sub-question.

Why the Same Question Gives Different Sources

Overviews change, though less than other engines. SISTRIX's 2026 study found AI Overviews far steadier from one week to the next than AI Mode or ChatGPT, but the average hides a split: most prompts never changed a source while a minority churned heavily. The figures are in how often AI answers change. Google's models also change. SE Ranking's 100,000-keyword study found about 42 percent of previously cited domains were replaced after Gemini 3 became the default model for AI Overviews on 27 January 2026, a sign that model changes move citations.

What You Can Do

  1. Check the basics in Search Console. Is the page indexed, eligible for snippets, free of an accidental nosnippet? Google's robots meta tag documentation says nosnippet prevents a page being used as direct input to AI Overviews and AI Mode, and max-snippet limits how much.
  2. Use Search Console's generative AI report. Google added a report of impressions from AI Overviews and AI Mode. It does not give clicks, position or queries, so treat it as a presence check.
  3. Write for the sub-questions. A page that answers the follow-up questions a buyer would ask has more routes into the fan-out. Evidence that this raises citation is observational, so treat it as sensible writing rather than a proven lever.
  4. Do not rely on rank alone. Check what an Overview actually cites for your questions. Track it over many answers, since a single view is one sample.
  5. Know the opt-out. A per-property Search Console control, available worldwide from 31 August 2026, can remove a site from AI Overviews and AI Mode without removing it from Search.

What Not to Do

Google's spam policies cover scaled content produced to manipulate Search, including its AI responses, and doorway pages written for similar queries. Publishing thousands of near-identical pages to cover every fan-out guess puts you in that territory. Its guidance also says seeking inauthentic mentions is less helpful than it might seem.

Lectern checks what Google AI Overviews say about your brand alongside ChatGPT, Gemini and Perplexity, on a fixed set of buyer questions, with the history kept. The free Hobby plan includes AI Overviews in its weekly check. See the pricing page for details.

Written by

Edgar Li

Edgar Li

Cofounder at Lectern

Edgar is a cofounder at Lectern, helping growth-stage companies teach AI models to accurately represent and recommend their products - turning that visibility into high-intent traffic and revenue. A product builder who thinks in narrative and customer value, he now applies that lens to helping founders win in AI search.