Query Fan-Out: What AI Search Really Does With Your Keyword

Query fan-out is the technique AI search systems use to break one typed question into a set of separate sub-queries, run them all at once, and assemble a single answer from whatever comes back. Google confirmed the mechanism when it launched AI Mode, and it has quietly rewritten the arithmetic of search visibility. Your rank tracker still shows one keyword. The engine is running a fan of searches you can’t see, and the citation goes to whoever covers the most of them.

This piece explains what query fan-out is, how many sub-queries the engines actually generate, what the 2026 citation data says about whether your #1 ranking still matters, how to find the fan-out queries in your own category, and what to change on the page once you have them. Every number below carries a named, dated source.

Nothing here is modeled or estimated.

What is query fan-out?

Query fan-out is the process by which an AI search system takes one user question, splits it into multiple related sub-queries, runs those searches in parallel, and synthesises the results into a single answer. Google’s own description of AI Mode is direct about it: the system “breaks down your question into subtopics and issues a multitude of queries simultaneously on your behalf,” which it says lets Search “dive deeper into the web than a traditional search.”

Robby Stein, Google’s VP of Product for Search, gave a plain-language version of the same thing in a July 2025 interview reported by Search Engine Journal. Ask “things to do in Nashville with a group,” he said, and the system “may think of a bunch of questions like great restaurants, great bars, things to do if you have kids, and it’ll start Googling basically.” Stein described AI Mode as using Google Search itself as a backend tool: issuing its own queries, then combining the results into one response with links.

The important part for marketers is what it does to the unit of measurement. For twenty-five years, one query meant one SERP and one ranking. Under fan-out, one query means a fan of retrievals, each with its own result set, and the answer you’re competing to appear in is stitched together from all of them. The user never sees those sub-queries. Neither does your rank tracker. Google’s December 2024 patent US12158907B1, “Thematic Search,” describes the same shape: a single query triggers sub-queries built from inferred themes, results are grouped by theme, and a language model writes a summary over the group.

How many sub-queries does one search actually generate?

Enough that the difference is structural rather than marginal, and the number scales with how hard the question is. Google has not published a fixed figure for standard AI Mode responses. What it has confirmed is the upper end: Stein said that when Google’s systems judge a query to need deeper reasoning, a feature called Deep Search takes over and can issue “dozens or even hundreds” of background queries, taking several minutes to finish rather than milliseconds.

The measured evidence comes from the Surfer SEO study published in December 2025. Working from 10,000 keywords, the researchers extracted 33,000 fan-out queries using Gemini, on a sample where 76% of the keywords triggered an AI Overview at all. Treat that as a floor for retrieval breadth, not a ceiling: it counts the fan-outs the study could reconstruct, not every query Google ran internally.

It also varies sharply by engine, which is where a lot of fan-out commentary overreaches. AirOps, working with Kevin Indig on a study published in April 2026, scraped ChatGPT’s interface across 16,851 unique queries and 353,799 pages and logged every sub-search the model issued. It found 88.6% of queries generated exactly two fan-out sub-queries, 8.8% generated none at all, and only 2.5% generated four or more. So “one question becomes hundreds” is a fair description of Google’s Deep Search and a poor one of a typical ChatGPT answer.

Check the engine before you size the opportunity.

The number you should care about is not the exact count anyway. It’s the volatility. The same Surfer research found that only about 27% of fan-out queries stayed consistent when the identical search was repeated: roughly three quarters changed run to run, shifting with phrasing, user context and personalisation.

That single finding governs the whole strategy, and we’ll come back to it.

Does ranking #1 still get you cited in AI Overviews?

Less reliably than it did a year ago. Ahrefs, analysing 863,000 keywords and 4 million AI Overview URLs, found that the share of AI Overview citations coming from pages ranked in Google’s top 10 for the same query fell from 76% in mid-2025 to 38% in early 2026. The rest split almost evenly: 31.2% of citations went to pages ranked between positions 11 and 100, and 31% to pages that didn’t rank in the top 100 for the query at all.

Ahrefs was careful to note that part of that drop reflects improved citation detection in its own tooling rather than a pure change in Google’s behaviour, so read the trend rather than the decimal. But the direction survives the caveat, and fan-out is the obvious mechanism. If the engine retrieves against a set of sub-queries and only one of them is the term you rank for, every other retrieval pass goes looking somewhere you’re not. A page that ranks nowhere for the head term can still be the best answer to one of the sub-queries, and that’s the one that gets quoted.

This is the same shift we traced in owning your search shelf-space beyond the ten blue links: position is no longer the currency, presence across the surface is. Fan-out just makes the surface bigger and less visible.

How much does ranking for fan-out queries actually help?

A great deal. It is the strongest lever in the published citation research. The Surfer SEO analysis of 10,000 keywords and 173,902 URLs found a strong correlation (Spearman 0.77) between the number of fan-out queries a page ranks for and its likelihood of being cited in an AI Overview. Pages ranking for fan-out queries were 161% more likely to be cited than pages ranking only for the main query.

The breakdown is where it gets useful. Pages that ranked for the main query and at least one fan-out accounted for 51% of AI Overview citations. Pages that ranked only for the main query accounted for just under 20%. When AI Overviews cited an organic result at all, roughly 20% of those citations went to pages ranking only for the head term while roughly 30% went to pages ranking only for fan-out queries. Ranking for the fan-outs was 49% more likely to earn a citation than ranking for the head term alone. The head term is now the smallest slice of the opportunity.

Independent evidence points the same way. In May 2026, Cyrus Shepard published a meta-analysis on Zyppy Signal distilling 54 experiments, patents and case studies into 23 AI citation ranking factors, each scored on repeatability across studies, strength of evidence, and official support from platform documentation or patents. “Fan-out rank” (how well a URL ranks for the related fan-out queries) scored 9.3 out of 10, third overall behind URL accessibility (9.5) and plain search rank (9.4), and ahead of every content-quality signal on the list. For scale, the lowest-scoring factor in that same analysis was llms.txt, at 2.0.

One important dissent, and it sharpens the advice rather than undoing it. The AirOps and Kevin Indig study of ChatGPT found that retrieval rank (where a page lands in the results ChatGPT gets back from its own sub-searches) was the single strongest predictor of citation: 58.4% citation rate at the first position against 14.2% at the tenth. On the same dataset, how many fan-out subtopics a page covered barely mattered, adding only 4.6 percentage points between zero coverage and full coverage. What did move it was how closely the page’s headings matched the query: pages with a close heading match were cited 41% of the time against 29% for weak matches.

So the honest reading across all three studies is not “cover everything.” It’s this: you have to rank for the sub-queries to be retrieved at all, and once retrieved, the page that gets quoted is the one whose headings visibly answer the question. Breadth gets you into the retrieval set. Precision gets you cited. Padding a page with shallow coverage of every subtopic does neither: AirOps found pages covering a moderate share of the fan-out outperformed pages that tried to cover all of it.

How do you find the fan-out queries for your topic?

You reconstruct them from the same signals the engine uses, then check your coverage against the list. This is a research workflow, not a tool purchase, and it runs in five steps.

1. Start with the buying question, not the keyword. Write the question a real prospect would type or say — “best CRM for a small agency,” not “crm software.” Fan-out is triggered by intent, and the more conversational the question, the wider the fan.
2. Decompose it the way a buyer would. List every sub-question that has to be answered before someone can act: pricing, alternatives, integrations, migration effort, contract terms, what goes wrong. Google’s own patent example splits “moving to Denver” into neighbourhoods, cost of living, and things to do. Your split should feel just as obvious.
3. Mine the question data. Pull the People Also Ask set and question-keyword report for your head term from a keyword tool. These are real, measured queries, and they overlap heavily with what fan-out generates because both are modelling the same underlying curiosity. Add the follow-up questions your sales team gets asked on calls — those rarely appear in keyword tools and almost always appear in fan-outs.
4. Add the modifiers. Commercial and temporal modifiers — best, top, reviews, alternatives, and the current year — get attached during fan-out. If your page never uses that language, it won’t match the modified sub-queries even when the substance is right.
5. Score your coverage, not your rank. For each sub-question, mark whether you have a passage that answers it directly — not a page that’s vaguely about it. Coverage percentage against the sub-question list is the metric. It is the closest thing to a fan-out rank you can measure without a platform.

One warning. Fan-out simulator tools have multiplied over the past year, and they can be a useful shortcut to a first draft of the list. What none of them can do is tell you which sub-queries Google actually ran, because the engine doesn’t publish them and, per the Surfer data, three quarters of them change between runs anyway. Treat any generated fan-out list as a hypothesis about the topic, not a ground truth about the SERP.

What should you change on the page once you know the fan-outs?

Widen the page’s coverage and structure it so each sub-question has its own liftable answer: then stop optimising for the individual fan-out strings. That second half matters as much as the first. With only about 27% of fan-out queries stable across repeated runs, building pages aimed at specific sub-query phrasings means aiming at a target that moves every time someone searches. Surfer’s own conclusion from its data was blunt: don’t chase fan-out queries, own the topic.

In practice, owning the topic looks like this:

• One question, one section, one direct answer. Lead each section with a self-contained one or two sentence answer, then expand beneath it. A retrieval pass on a sub-query is scanning for a passage it can quote whole. This is exactly the discipline that wins featured snippets and People Also Ask placements, pointed at a new consumer.
• Headings phrased as the questions themselves. Match the language of the sub-question rather than a clever label. The match between the query and your heading is doing retrieval work.
• Enough breadth to be retrieved, not breadth for its own sake. Six real answers to the six sub-questions that matter will beat one beautiful answer to the head term — and also beat a bloated page that gestures at thirty. Moderate, precise coverage outperformed exhaustive coverage in the AirOps data.
• Spread across the cluster, not one mega-page. Related pages that each own a narrow sub-question give the engine more distinct retrieval targets than a single page trying to hold everything.
• Keep the modifiers and dates natural. Comparison language, review framing and current-year references belong in the copy because the sub-queries carry them — not as keyword stuffing, as the way people actually ask.

None of that is a new discipline. It is structured, answer-first content (the same work that has been winning search features for a decade) scaled sideways across a topic instead of concentrated on one term. What’s changed is the payoff curve: under the old model the second-best answer on a page earned nothing, and under fan-out it might be the passage that gets quoted. For the mechanics of how engines pick between the passages they retrieve, we broke that down in how LLMs actually choose their sources, and the wider discipline sits inside our SEO practice.