Fan-Out Queries: Definition, and Why They Decide Your AI Visibility

Since Google I/O in May 2025, the technique has had an official name and an official description. In its announcement of AI Mode, Google wrote that under the hood, AI Mode uses its query fan-out technique, "breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf". Read that sentence slowly, because it quietly retires an assumption the whole SEO industry was built on: that the query a user types is the query the engine runs. It is not, anymore. This page defines fan-out precisely, shows how the loop works in production, and explains why the queries you can no longer see are the ones now deciding whether you get cited.

Fan-out queries, defined in one line

Fan-out queries are the sub-queries an AI search system generates from a single user question, then runs in parallel to build one synthesized answer. Google calls the mechanism its "query fan-out technique": it breaks your question into subtopics and issues a multitude of searches simultaneously on your behalf.

The definition is Google's own, not mine, and that matters. Its Search Central documentation on AI features states that both AI Overviews and AI Mode "may use a 'query fan-out' technique, issuing multiple related searches across subtopics and data sources, to develop a response". Two separate Google surfaces, one shared mechanism, described in Google's own words.

The name is a good one, for once. Picture a beam of light hitting a prism: one thing goes in, many things come out, spread across a surface. Your question is the beam. The subtopics are the spectrum. Somewhere in that spread, a page gets picked to answer each band, and that page either belongs to you or it does not.

One scoping note, because this vocabulary is a swamp. Fan-out is a retrieval mechanism inside the engine. The disciplines built on top of it carry other names: GEO for earning citations in generated answers, AEO for being served as the direct answer. And it runs on the surfaces you already know, AI Overviews and AI Mode. This page stays on the mechanism.

How the fan-out actually runs

Three steps: decomposition (the model splits the question into subtopics), parallel retrieval (it issues many searches at once, across sources), and synthesis with attribution (it writes one answer and links the pages it leaned on). Google says Deep Search runs the same technique further, issuing hundreds of searches for a single report.

Here is the loop, in the order the machine executes it.

Scale matters here. Google describes Deep Search in AI Mode as using "the same query fan-out technique but taken to the next level", able to "issue hundreds of searches, reason across disparate pieces of information, and create an expert-level fully-cited report in just minutes". Hundreds of searches, for one question. Each one is a door someone walks through, and doors are what you are optimizing for now.

None of this appeared out of nowhere. Decomposing a hard question into answerable sub-questions is a documented retrieval technique with a research trail behind it. Jagerman and colleagues at Google Research showed in their 2023 paper on query expansion by prompting large language models that chain-of-thought prompts are especially useful for query expansion, because they instruct the model to break queries down step by step and yield a large number of related terms. Press and colleagues, in their work on the compositionality gap, found that models answer multi-hop questions far better when made to state and answer the sub-questions explicitly. Fan-out is that lab result, shipped to a billion people.

Nor is it Google's alone. OpenAI's web search tool returns inline citations, each cited page coming back as a url_citation annotation carrying the URL and title. Different vendor, different naming, same structural consequence: the engine searches on your behalf, then names the pages it used.

The keyword you optimized for is never run

A keyword is what a human types. A fan-out query is what the machine runs next, and you never see it. Your page is not retrieved for the user's question, it is retrieved for one of the sub-queries derived from it. That is how a page can dominate a head term and be absent from the AI answer to that same term.

This is the part that should reorganize a content plan, so let me be blunt about it. In a fan-out world, the keyword is an input to a machine that discards it. What survives is the set of subtopics the model decided the question contained. Nobody shows you that set. It varies with phrasing, with the model, with the day.

The consequence is counter-intuitive and, I think, under-appreciated: fan-out is good news for smaller sites. Google says so in the same breath, noting that while responses are generated, its models identify more supporting web pages, which lets it show "a wider and more diverse set of helpful links associated with the response than with a classic web search". More retrieval passes mean more slots. A specialist page that answers one narrow subtopic better than anyone can be pulled into an answer whose head term it would never have ranked for.

Pew Research Center's July 2025 analysis puts a number on the width: 88% of the AI summaries it examined cited three or more sources. Three or more, per answer. The room is not crowded because it is small, it is crowded because most pages are not built to fit through any of the doors.

Fan-out lives in the long tail

Sub-queries are narrower than the question that spawned them: price, method, comparison, exceptions, edge cases. Those are exactly the queries keyword tools report as too small to matter. Fan-out converts that discarded demand into the retrieval surface, which is why coverage now beats volume.

Follow the logic and it becomes uncomfortable for anyone who plans content off a volume column. If the engine decomposes "which insurance should I take for a rented flat" into ten sub-questions about coverage, exclusions, deposits and legal minimums, then those ten are what get searched. Look them up in a keyword tool and most of them will report a shrug.

Our own corpus says the shrug is the norm, not the exception. According to Cicero Studio's internal data, across the 4253 French keywords Cicero Studio measured, the median volume is 260 searches/month, and 34% of the French keywords Cicero Studio analyzed get fewer than 100 monthly searches: the long tail dominates. A third of real demand sits below the threshold at which most strategies even look. Fan-out does not care about that threshold. It searches the sub-question because the sub-question is what the answer needs.

Pew's data points the same way from the user side. It found that 60% of search queries beginning with question words such as who, what, when or why produced an AI summary, against just 8% of one-word or two-word searches. In other words, the queries that trigger the AI machinery are precisely the long, question-shaped ones, and the short head terms that keyword strategy fetishises mostly do not trigger it at all. Two independent datasets, one conclusion: the action moved to the tail.

How to be visible on queries you cannot see

You cannot target invisible sub-queries, so you make each one independently answerable: one question per heading, a self-contained answer directly beneath it, a named source behind every claim, structured data so machines parse rather than guess, and internal links so no answer sits orphaned.

None of it is exotic. It is mostly what happens when you write honestly, then make that honesty machine-legible.

Here is a free diagnostic you can run this afternoon, no tool required. Take a real customer question. Write down, honestly, the ten sub-questions someone would have to settle before answering it. Now search your own site for each. The gaps you find are, near enough, the fan-out queries you are already losing. In my experience the exercise is unpleasant, because the missing answers are almost never the exotic ones. They are the obvious ones nobody bothered to write down.

Across the 1207 SEO/GEO audits Cicero Studio has produced, the pattern repeats to the point of monotony. The page that gets cited is rarely the flagship. It is some short, unglamorous FAQ or definition page that happens to settle one sub-question in one sentence, while the flagship carries more genuine expertise and takes nine paragraphs to reach its point. Nothing about authority explains that outcome. Format does. Which is why our most common fix is not a new article but moving the answer from paragraph six to paragraph one, naming the source that was already implicit, and marking the page up so the machine stops guessing.

What fan-out does not change

Now the honest part, which matters more than usual here, because fan-out is currently being sold as a secret door.

What fan-out does give you is a mechanism you can work with, which is rarer than it sounds. The engine has to search something, and it has to cite something. Deciding to be the page that answers the sub-question is a strategy. Waiting to be discovered is not.

Alexis Dollé

CEO & Founder of Cicero Studio

A growth specialist and content strategy consultant, I founded Cicero to help businesses build durable organic visibility, on Google as in AI answers. Day to day, I run our clients' audits and editorial production: we put AI to work for production, never in place of expertise. Every piece is built to convert, not just to exist.

LinkedIn →

Where Cicero Studio fits

Cicero Studio treats the question space, not the keyword, as the unit of work: a GEO audit that measures which sub-questions you can realistically own, editorial production built around liftable sourced answers, and automated semantic internal linking so no answer stays orphaned. It starts with a free audit.

The method hook is easy to say and harder to run: GEO audit, then editorial production, then automated semantic meshing, as one loop rather than three disconnected services. Fan-out is what ties them together. The audit maps the sub-questions a retriever could plausibly pull you into. The production makes each answer worth lifting. The meshing makes sure a crawler reaches it at all.

We run the same experiment on ourselves, in the open, across the 502 articles published on cicero.studio (264 FR, 238 EN), each written to be extractable rather than merely readable. That is what we mean by agency-quality work, software-grade productivity. The French treatment of the model lives on our agence GEO pillar, and this page has a French sibling: fan-out queries, la définition.

Going further

Fan-out is the mechanism. The surfaces it feeds and the disciplines built on top of it have their own names, and keeping them straight is most of the battle. Read them in this order and the alphabet soup resolves: fan-out is how the machine searches, grounding is how it ties the answer to sources, passage ranking is how it picks which piece of your page to use, and GEO is what you do about all three.

We publish our method in the open rather than in a sales deck, for the same reason this page quotes Google's documentation instead of asking you to take our word for it. Several pieces below are in French, our home market, and are flagged as such.