Share of answer is the operating metric for this practice.

How we count it
Request an audit

Insight · 11 min

How AI Overviews choose a source

Google AI Overviews, the extraction surface in SEO, AEO, and GEO, cite pages that Google’s core ranking systems retrieve from the Search index, both for the query the person typed and for a set of related sub-queries the model generates, a process Google calls query fan-out. To be cited, a page has to be indexed, eligible for a snippet, and the best available answer to one of those sub-questions. Ranking on page one for the exact query helps, but as of 2026 it no longer predicts the citation: in Ahrefs’ March 2026 study, only 38% of cited pages ranked in the top 10 for the same query.

In short

  • Google’s only technical requirements are an indexed page that is eligible for a snippet, on a site that has not opted out of generative AI features. There is no markup, file, or format that opts you in.
  • AI Overviews are grounded in Search. Core ranking systems retrieve pages, and the model writes the answer from them and links the pages that support it.
  • Fan-out is why the citation and the ranking have come apart. Ahrefs found 38% of AI Overview citations in the top 10 in March 2026, down from 76% in July 2025, with a changed method.
  • Win the sub-questions, not more URLs. One page that answers the main question and its likely follow-ups, under clear headings, beats a page per variant. Google calls the second approach scaled content abuse.
  • A citation is exposure, not traffic. In Pew’s 2025 study, people clicked a link inside the AI summary on 1% of visits. Measure impressions and accuracy, and treat clicks as a separate number.

Google picks them by retrieval first and generation second. Its guide to optimizing for generative AI features describes the system as retrieval-augmented generation, also called grounding: Google’s core ranking systems retrieve pages from the Search index, the model reviews those pages to write the response, and the response shows “prominent, clickable links to relevant web pages that support the information in the response.”

Fan-out changes which pages are in that retrieved set, and other engines do it too, as the guide to query fan-out compares. Google defines it as “a set of concurrent, related queries generated by the model to request more information.” Its own example: a search for how to fix a lawn full of weeds might also issue queries about herbicides, chemical-free removal, and prevention. The “AI features and your website” page says fan-out lets Google “display a wider and more diverse set of helpful links” than a classic result, and that it identifies more supporting pages while the response is being generated.

Three more facts from Google’s documentation shape the target:

  • AI Overviews appear only when Google’s systems judge them additive to classic results, so many queries have none.
  • AI Overviews and AI Mode may use different models and techniques, so the same query can cite different links in each.
  • Google moved AI Overviews to Gemini 3 as its default model worldwide in late January 2026. Model changes can shift which sources are chosen without any change on your site.

So the selection is not one ranking. It is a set of rankings, one per sub-query, followed by a model deciding which retrieved passages support which sentence of its answer.

Do you need to rank in the top 10 to be cited?

No, and the share of citations that come from the top 10 has fallen sharply. Ahrefs’ update of March 2, 2026 matched about 4 million AI Overview URLs against the regular results for 863,000 keywords. Of the cited pages:

Where the cited page ranked for the same queryShare of citations
Top 1037.9%
Positions 11 to 10031.2%
Not in the top 10031.0%

Restricted to standard blue links only, 37.1% were in the top 10 and 36.7% were outside the top 100. Ahrefs’ July 2025 version of the study put the top-10 share at about 76%.

Read the number with its caveats. Ahrefs changed its parsing between the two studies, so the drop is not like-for-like. The data comes from one vendor’s tool. Ahrefs’ explanation, that AI Overviews now lean on the pages that appear most often across fan-out results, is a hypothesis: Google does not publish its fan-out queries. Search Engine Journal reported a February 2026 BrightEdge figure near 17%, which we could not find at a BrightEdge primary source. Other studies use other methods and land elsewhere; the AI search statistics reference lists each one with its sample and denominator.

The practical reading: a page can be cited for a query it does not rank for, because it ranks for a question the model asked on the way to the answer. A page can also rank first and not be cited, because a narrower page answers a sub-question better. Rank tracking on the head query alone will miss both.

What does Google say you need to do?

Google says you need nothing beyond normal Search eligibility, and lists what not to bother with. The requirements are a page that is indexed and eligible to show with a snippet, on a site included in generative AI features in Search Console. Google adds that meeting them does not guarantee crawling, indexing, or serving.

What Google recommends, from its May 2025 post by John Mueller and its generative AI features guide:

  • Unique, “non-commodity” content, written from what you know, including first-hand experience, rather than a restatement of what already ranks.
  • Paragraphs, sections, and “headings that provide a clear structure to navigate content.”
  • Important content in textual form, with crawlable pages, semantic HTML where practical, and sound JavaScript rendering. Most non-Google AI crawlers do not run JavaScript at all.
  • A page that makes it easy to tell main content from everything else.
  • Images and video where they help, and current Merchant Center and Business Profile data for products and locations.

What Google says you do not need:

  • Chunking. “There’s no requirement to break your content into tiny pieces for AI to better understand it.” Google can find the relevant part of a long page.
  • AI files or special markup. No llms.txt, no AI text files, no Markdown copies. Google Search ignores them. The llms.txt piece covers who does read them.
  • Special schema. Structured data still earns rich results, but it is not an AI Overviews switch. See what schema does and does not do.
  • Writing for the machine. The systems understand synonyms, so chasing every long-tail variant is unnecessary.
  • Manufactured mentions. Seeking artificial mentions across the web “isn’t as helpful as it might seem.”

One warning matters more than the rest. Google says that creating pages for every fan-out variation to manipulate AI responses violates its scaled content abuse spam policy. Fan-out is a reason to answer the follow-up questions well, not a reason to publish a URL per question.

How do you build a page that gets cited?

Build one canonical page per intent that answers the main question and the questions a model is likely to ask next, each one in a passage that can stand alone. Google does not publish a recipe beyond the guidance above. What follows is how we apply it in an AEO engagement, and it is a method, not a guarantee.

  1. Write the sub-questions down. Start from the query universe, not from a prompt generator. For each priority query, list the follow-ups a buyer would ask: cost, time, conditions, alternatives, risks, who it is not for. The People Also Ask box, sales calls, and support tickets are better evidence than a tool’s guesses.
  2. Decide which page owns each one. Most belong on the main page as a section. A sub-question earns its own URL only when its intent is distinct and nothing on the site serves it, which is a mapping decision, not a fan-out decision.
  3. Answer in the first sentence under the heading. Phrase the heading as the question. Make the first sentence the answer, with the number, condition, or distinction in it. Put the evidence in the next sentence: who measured it, and when.
  4. Keep the claim in HTML. Prices in a PDF, specs in an image, and hours in an embed are invisible to the passage the model needs. The Kelso Motion sample readout is that failure on spec pages.
  5. Make the claim checkable elsewhere. Engines prefer claims that independent sources agree with. That is the corroboration work described in what gets cited.
  6. Fix retrieval before passages. A page with competing near-duplicates, a wrong canonical, or blocked rendering cannot be retrieved for any sub-query. That is SEO, and it comes first.

A self-contained passage is not the same as chunking. The page stays a normal page written for a reader. Each section simply answers its heading without depending on the paragraph above it, which helps a skimming person as much as a model.

Which kinds of sources do AI Overviews cite?

AI Overviews usually cite several sources, and a few large platforms take a big share. Pew Research Center analyzed 68,879 Google searches made by 900 U.S. adults in March 2025, of which 12,593 produced an AI summary:

  • 88% of summaries cited three or more sources. 1% cited a single source.
  • Wikipedia, YouTube, and Reddit together were 15% of the sources cited in AI summaries, and 17% of the links in standard results.
  • Government sites were 6% of AI summary sources, against 2% of standard result links.
  • The median summary was 67 words.
  • Summaries appeared on 8% of one- and two-word searches, 53% of searches of ten words or more, and 60% of searches that began with a question word.

Ahrefs’ 2026 data points the same way for video: YouTube URLs were 5.6% of all AI Overview citations, and 18.2% of the citations that came from outside the top 100.

Two implications. Longer, question-shaped queries are where AI Overviews appear, so that is where the passage work pays. And the citation slot is shared: you are competing to be one of several sources, often next to a forum thread or a video. A clear explanation on your own site, plus the same facts stated consistently on the platforms the engine already trusts, is the realistic goal.

Is an AI Overview citation worth the work?

It is worth it as visibility and as control of how you are described, and it is a weak traffic source. Pew found that on visits where an AI summary appeared, users clicked a traditional result 8% of the time, against 15% without one, and clicked a link inside the summary on 1% of visits. Users ended their browsing session on 26% of visits with a summary, against 16% without.

Ahrefs’ February 2026 update estimated that position-one click-through rate was about 58% lower than forecast for keywords with an AI Overview, using desktop Search Console data from December 2023 and December 2025 across 300,000 keywords. That figure is an estimate built on a forecast, and Ahrefs describes it as a correlation. The CTR studies disagree with each other for reasons of method, set out in AI search statistics, compared, and diagnosing click loss from AI Overviews shows how to check your own Search Console data. Google’s position is that clicks from pages with AI Overviews are “higher quality,” with users more likely to spend more time on the site. It has not published numbers that let anyone check that.

The answer still gets read when nobody clicks. If the overview names a rival, quotes a retired price, or describes your product wrongly, the buyer acts on that, which is why finding and fixing what AI gets wrong about your brand is part of the work. Score presence and accuracy, as described in share of answer, and report clicks as their own number.

How do you measure AI Overview citations?

Use Search Console for scale and a logged query panel for content, because neither shows both. In the standard Performance report, AI Overview and AI Mode traffic is counted inside the “Web” search type and is not separated from other results. All links in one AI Overview share a single position.

The Generative AI performance report, announced June 3, 2026 and rolled out to all sites by August 31, 2026, shows impressions of your URLs in AI Overviews and AI Mode, by page, country, device, and date. It does not show queries or clicks. It tells you which pages are appearing. It does not tell you for what, or what the answer said.

For that, run the same reviewed queries on a fixed schedule, repeat each one, and log the overview, its sources, and the date. Google also warns that “no third-party tool has access to our internal ranking or AI systems,” so treat any tool that promises otherwise with suspicion. The full method is in measuring AI search visibility.

Can you keep content out of AI Overviews?

Yes, with Search controls, and each one costs something in classic results too. Google lists nosnippet, data-nosnippet, max-snippet, and noindex as the ways to limit what Search shows, including in AI features. data-nosnippet is the narrowest: it hides one element, such as a price table, from snippets and AI Overviews while the rest of the page stays eligible. Search Console also offers a Search generative AI control that excludes a site from these features entirely.

Two controls do not work. Google-Extended governs training and grounding for Gemini Apps and Vertex AI, not AI Overviews. Blocking Googlebot removes you from Google Search altogether. The robots.txt guide covers both.

What are the common mistakes?

The common mistakes treat AI Overviews as a separate channel instead of an extension of Search.

  • Tracking only the head query. The citation may come from a sub-query you never measured.
  • A page per fan-out question. Thin near-duplicates compete with each other and risk the scaled content policy.
  • Buying an “AI Overviews schema” or llms.txt package. Google says neither is needed.
  • Burying the answer. A long introduction before the claim forces the model to find it elsewhere.
  • Facts only in PDFs, images, or scripts. If it is not in the HTML text, it is not available to quote.
  • Counting citations as traffic. One overview can cite you and send nobody. Report them apart.

AI Overviews reward the same thing featured snippets did: the clearest checkable answer to a specific question, on a page Google can retrieve. The change is that the questions now multiply behind the query, and the page that answers the follow-up can win the citation.

Sources

Bring us the domain.

We will tell you where you are absent, which page is in the way, and what has to be true before a citation is possible.

Request an audit