How Do Google's AI Overviews Decide What to Cite?

Updated 19 October 2026 · 10 min read · By Matthew Bertram

So a page doesn't only compete for the query someone typed. It competes across a cluster of follow-up searches the reader never sees, and a page that answers more of that cluster gives the system more chances to pick it.

What is query fan-out?

Google's AI optimization guide defines query fan-out as "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query." Its own example starts from "how to fix a lawn that's full of weeds" and fans out to "best herbicides for lawns," "remove weeds without chemicals" and "how to prevent weeds in lawn."

None of those searches repeats the original question. Each one chases a single part of the problem. A page about pulling weeds without chemicals could be cited under the lawn question without ranking for the lawn question at all.

Google's AI features documentation says AI Overviews and AI Mode "may use" the technique. It doesn't say how often, or for which questions.

This guide covers Google only. ChatGPT and Perplexity retrieve sources their own way, and our guide to answer engine optimization explains how the engines differ. Whether that means separate work for each engine is a question our guide to AI SEO answers.

How many searches does one question trigger?

Google hasn't published a number for AI Overviews. The one figure it has published belongs to a different product. Deep Search, a research mode inside AI Mode, "can issue hundreds of searches," Google said in May 2025. Quoting that number for AI Overviews mixes up two products.

The best outside measurement we found comes from Seer Interactive. In November 2025 it ran 501 prompts through the Gemini 3 API with Google Search grounding forced on and got back an average of 10.7 searches per prompt, with a low of 3 and a high of 28. Seer puts the Gemini 2.5 average at 6.01. That's Google's model reached through its developer API, not AI Overviews themselves, and it's one study. Treat it as a proxy, though a closer one than it sounds: Ahrefs reports that AI Overviews have run on Gemini 3 since January 2026.

One more Seer finding changes how you research a topic. 95% of the searches Gemini wrote had zero global search volume. Seer is careful to say that doesn't mean nobody wants the answer. It does mean a keyword tool that starts from volume will miss most of the set.

The mechanism also appears in a Google patent filing, with a caveat. A pending application, "Search with stateful chat" (US20240289407A1, published August 2024), describes a generative model that takes the user's query plus contextual information about the user or their device and produces "synthetic queries." The system then selects search results that respond to both the original query and the synthetic ones. An application shows what Google's engineers filed. It doesn't prove what runs inside AI Overviews today.

What does a fan-out set look like?

Take a question our guide to blocking AI crawlers answers: "should I block AI crawlers in robots.txt." Here is what a fan-out set for it plausibly contains.

Illustration, not a capture: Google doesn't show the searches behind an AI Overview, so these rows are a reconstruction. We built them from People Also Ask and related searches. One row adds a year and one adds a vendor, because in Seer's data 21.3% of Gemini's fan-outs contained a year and 26.4% named a brand. Trust the shape of the table more than any single row.

Each row is a separate question with its own answer. A page that answers only the first question competes for one search in that set. A page that gives each row its own heading and a direct answer competes for most of them. That's the whole argument for how these guides are built: one topic per page, one question per heading. The same holds outside crawlers: GEO versus AEO is a "telling two things apart" search, so it has a guide of its own.

Can I see the real fan-out queries for my topic?

Not from Google. Search Console's generative AI performance report, which Google says reached all websites on August 31, 2026, counts AI Overview and AI Mode impressions by page, country, date and device. It has no query view, so it can't show you the searches behind an answer. Surfer, which studied this, says it plainly: "Right now, there's no way to see exactly which fan-out queries AIOs actually search for." Its researchers used Gemini as a stand-in.

Microsoft shows more than Google does. Bing Webmaster Tools' AI Performance report lists "grounding queries," which Microsoft describes as "the key phrases the AI used when retrieving content that was referenced in AI-generated answers." It's a sample, and it covers Microsoft's AI answers, not Google's. It's still the only report we know of that ties retrieval searches to your own site's citations, and our guide to checking whether AI engines cite you walks through it.

For Google, you approximate from four sources:

1. People Also Ask on the head question, opened two levels deep.
2. Related searches at the bottom of the results page.
3. The follow-ups chat engines suggest after they answer. That's the engine showing you what it expects you to ask next.
4. Your own Search Console queries for pages that already rank on the topic. Those are real phrasings from real people.

Merge them and drop the duplicates. You won't have Google's exact set, and you don't need it.

Does ranking first still get you cited?

It helps. It doesn't decide. Every large study we found has a big share of AI Overview citations going to pages outside the top 10, but the studies counted different things and disagree about how big that share is.

Method explains part of the spread. Ahrefs' 2025 study counted only the three most visible citations in each answer, and its 2026 update used an improved parser that sees more of them, so the drop from 76% to 38% isn't purely a change in Google. The update's author reads the drop as a sign that AI Overviews now lean more on pages found through fan-out searches. Google hasn't confirmed that. Surfer checked only the top 10 for each search and says itself that many of the other cited pages may simply rank lower.

The direction is disputed too, though the two studies measure different things. seoClarity's repeat analysis found 3 top-20 URLs in the average AI Overview in October 2025, down from 5 in May. BrightEdge, tracking an earlier window that ended in September 2025, saw overlap with organic rankings rise from 32.3% to 54.5%.

Comparisons made inside one study are sturdier, because the method is the same on both sides. Two of them matter here. In seoClarity's data a URL in position one was cited 43% of the time, falling to 7% by position 20. And when Surfer counted only the top three citations, the ones visible without clicking "Show all," 54% ranked in the top 10 for the query or one of its fan-outs, against about 32% for all citations.

Coverage is where fan-out shows up. In Surfer's 173,902-URL study, the more fan-out searches a page ranked for, the higher its odds of being cited, with a Spearman correlation of 0.77. Among cited pages that ranked at all, 51.2% ranked for the main query and at least one fan-out, 29.2% for fan-outs only and 19.6% for the main query alone. Surfer says itself that this is correlation, not proof of cause. It still points one way. A page built around one exact phrase is entered in one search out of a set.

Why did my AI Overview citation disappear?

Usually because the set of candidate pages moved, not because you were penalized. The fan-out searches themselves change from run to run. Surfer found that only about 27% of Gemini's fan-outs stayed the same across repeat searches for the same prompt. They also change with the person asking. Surfer's illustration: for "best electric car," a searcher Google knows has a family gets a fan-out like "electric car safety ratings," and someone else gets "fastest electric cars." The patent application above lists the same input, contextual information about the user and their device.

Our own variance study didn't include AI Overviews, so we can't give you an AI Overview figure. Across the five answer engines it did cover, the finding was blunt: "A month of apparent movement is statistically indistinguishable from asking twice." One missing citation on one check tells you almost nothing. A drop that holds across repeated checks of the same questions does, and our guide to checking whether AI engines cite you covers how many repeats that takes.

If the drop is real and sustained, check the boring causes first. Is the page still indexed? Did a template change add a nosnippet rule? Did someone edit robots.txt? Google's on-record position is that a page has to be indexed and eligible to show with a snippet before an AI Overview can link to it. Then compare coverage. Look for a competitor that published something answering more of the cluster than you do, and for sub-questions you stopped answering when the page was edited.

How do you write a page for fan-out?

Write one thorough page per topic, and give every sub-question its own heading with an answer that stands alone. Most of the work is in the structure, not the sentences.

Run this check before a page goes live:

1. List the follow-up questions a reader would ask right after your first answer. Build the list from the four sources above, not from memory.
2. Answer each one outright. "It's covered in there somewhere" fails.
3. Make each answer stand alone. Cover the rest of the page and read the two sentences under a heading. If they don't answer it by themselves, rewrite them.
4. Give each answer its own heading, worded the way someone would type the question. A heading marks where one answer stops and the next starts, for a person skimming and for any system that splits pages into passages.
5. Don't defer a payoff. No "more on this below." A passage lifted out of the page takes none of the page with it.

A question that fails steps 2 to 5 is one you aren't competing for.

Should I write one long page or several short ones?

One page when the sub-questions share an intent. Separate, linked pages when they don't. The test is whether someone asking the sixth question on your list would be happy to land on the page that answers the first. If they would, it's one page.

Google's on-record position points the same way. Its AI optimization guide warns that creating "separate content for every possible variation of how people might search," including fan-out queries, "primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy." A separate page per fan-out search is the tactic Google names.

Even the research that found the fan-out correlation argues against chasing the searches one by one. Surfer's author considered the obvious next move, to "scrape as many fanouts as possible and try to rank for them all," and wrote: "This probably isn't the best idea." The searches shift between runs and between people, so a page aimed at one of them chases a moving target. Write for the question cluster a reader actually has, and give a section its own page only once it draws searches of its own.

Want to see which follow-up questions around your core topics your site answers, and which ones send Google's AI to a competitor? Book a discovery call with EWR Digital →

Keep reading: What Is Answer Engine Optimization (AEO)? · How Do I Know If ChatGPT or Google AI Overviews Are Citing My Site? · Should I Block AI Crawlers in Robots.txt?

Have you seen a real fan-out set for your own topic, in Bing's grounding queries or a tool's logs? Send it to us. We'll test it against the example in this guide and correct the guide where it's wrong.