Query Fan out

What it means in practice

Query fan-out is what an AI answer does before it writes anything. A person types a search. The model behind the answer does not simply find the best page for that exact phrase and summarise it. It breaks the phrase into several related sub-questions, looks for a source for each one, and builds the answer from what it finds. One search becomes several small lookups.

The consequence for a business is that the question typed into the box is no longer the only question being answered. A page can lose the main query and still be cited through one of the sub-questions. That is a different target from the one conventional keyword ranking was built for, and it is why this short term now sits underneath most of what the site says about AI Overviews and about being cited, not just ranked.

How it works

The ranking article gives the sequence plainly. Before answering, a large language model silently generates several variations of the question a person actually typed. It pulls candidate sources for each variation separately. It then merges the results with a technique such as Reciprocal Rank Fusion, which favours a source that shows up consistently across the set over one that dominates just a single version of the query.

The word “silently” matters. Nobody sees the variations. The person sees one search and one answer, and the sub-questions exist only inside the system. That is why fan-out is hard to observe directly and easy to underestimate, and why the practical advice later on this page is about testing from the outside.

Where it sits in the retrieval picture

Fan-out is not a separate technology. It sits on top of retrieval-augmented generation, the process of pulling a live, specific source into an answer instead of relying only on what the model learned during training. The ranking article puts it in one line: fan-out is that process run several times per search instead of once.

This also connects to grounding. The AI Overviews article says a properly grounded answer points back to a real, checkable source for each claim, while one that is not can read as confident and still be wrong. Fan-out multiplies the number of places a source gets chosen, so it multiplies the number of claims that need to be traceable.

A worked example

The ranking article offers its own illustration. A search for “AI training for HR teams” might fan out into related questions such as what AI bias training for recruitment covers, what the EU AI Act requires of HR specifically, and how CV screening tools get evaluated.

A page built narrowly around the first phrase, and nothing else, wins that one query if it ranks first for it. A page that covers all four angles with real depth, even if it only ranks sixth or seventh for the original phrase, gets pulled into more of the merged results, because it keeps showing up across the set the fan-out retrieves from. The article is explicit about the trade: depth across a topic now competes directly with dominance on one exact phrase.

A second, smaller illustration, invented for this page and not a real client. A person asks an assistant who can train a sales team to use AI. Behind the scenes the system might also ask what a short course for sales staff should cover, what a team should know about data before using a chatbot, and how a business checks that the training worked. A page that answers only the second of those questions, clearly and with a source beside each claim, can be cited for that part of the answer even though it never mentions sales training in its title.

Why the Citation figures moved

The ranking article uses fan-out to explain a large change in a short time. In a July 2025 study, Google’s AI Overview cited a top-ten page 76% of the time. A March 2026 follow-up on 863,000 keywords and 4 million AI Overview URLs put that at 37.9%. Roughly six in ten citations now come from a page that would not have appeared on the first page of ordinary results for the same search.

Ahrefs points to Google’s switch to Gemini 3 in January 2026 as the likely reason. The article says a newer model appears to run a heavier fan-out before it answers, pulling sources across more variations of a query than the model it replaced. The wording is cautious, and this page keeps it that way: it is a likely explanation offered by Ahrefs, not something Google has confirmed in the site’s sources.

The standalone assistants make the same point from the other side. In the same 2025 data, only 12% of ChatGPT, Gemini and Copilot citations came from a page ranking in Google’s top ten for the matching query, and Perplexity sat at 28.6%. The article reads this as the AI Overview moving towards a behaviour the assistants already had. It describes the earlier AI Overview as leaning on something closer to a single retrieval built from the query as typed, which is why it tracked ordinary rankings so closely for so long.

What it rewards

Two things follow, and the site’s articles state both. The first is depth on a narrow point over breadth attempted across a broad one. A page that tries to cover everything gives a model no reason to pick it for any particular sub-question, while a page that answers one thing clearly gives it a reason to pick it for that thing.

The second is that long-tail questions deserve more deliberate attention. The AI Overviews article says specific questions, ones the model has not already synthesised an answer to from elsewhere, are worth more attention than they were when a single head-term ranking was the whole game. Fan-out is what turns a long-tail question from a minor source of traffic into a place where a page can be named.

What still decides whether a page is in the running

Fan-out does not replace the usual gates. The AI Overviews article says authority signals matter more under this model, not less, because being selected as a trustworthy source is a precondition for being cited at all. Clear authorship, evidence of direct experience and a track record the model can find corroborated elsewhere are the signals that put a page in the running. The E-E-A-T page covers them.

The second gate is extraction. A model choosing what to cite is lifting a claim, not composing one. The ranking article names three structural habits with a measured effect: opening paragraphs that answer the query immediately are cited 67% more often, original data tables earn 4.1 times more citations than the same claim in prose, and FAQ schema is linked to a 28% rise in citation rate. It adds a CXL finding that 55% of cited passages came from the first 30% of the source page. These are reported figures from the article’s own sources, not measurements by this site. The generative engine optimisation page covers the wider set of levers.

Fan-Out and Ordinary SEO

Nothing here argues for dropping conventional search engine optimisation. The ranking article says the AI Overview still correlates with rankings more than the standalone assistants do, even after the drop from 76% to 37.9%, so the work that earns a page one still buys real value. What changes is treating that work as the whole job.

The article puts the new question this way: not only where a page sits on a results page, but whether a model reading across the top sources would choose it as one worth naming. A page structured to rank and a page structured to be cited want largely the same things, which are a direct answer, a clear structure and evidence stated specifically. They are simply checked against a different question.

How to see it from the outside

You cannot read the list of sub-questions a system generates, so the useful checks are indirect. The ranking article suggests looking by hand, and periodically, at what an AI Overview or an assistant says for the handful of queries that matter most, and which sources it names while saying it. A page that used to rank first can find a competitor cited in its place, or find itself cited only for a narrower sub-question than the one it was written to answer. Neither shows up in a rank tracker.

The brand article adds a method. It describes a fixed set of ten to twenty-five prompts, written the way a person would type them and not the way a marketer would, and warns that an assistant often splits one question into several searches. A category question can therefore surface sources you would never have searched for yourself. Read the sources named, not only whether your business appears, and the sub-questions start to show.

What to track

Two splits from the articles are worth building into reporting. The first is rank position against citation frequency. The ranking article says rank tracking alone no longer says whether a page is cited, on Google or inside an assistant, and that citation frequency needs its own tracking. It names Am I Cited, a tool this business partners with, as one way to do that over time.

The second is branded against non-branded queries, tracked as two lines and not one blended total. The article says a page can lose ground on generic, category-level searches, the ones most exposed to a fan-out picking a different source each time, while holding or gaining on searches for the business by name. A branded search is a query a generated answer has far less reason to reroute. The AI Overviews article adds a third signal: in Search Console, impressions holding steady or climbing while clicks on the same query drift down, which reflects a query being answered on the results page itself.

Mix-ups worth avoiding

Four confusions come up often enough to name. The first is treating fan-out as something a business can switch on or off. It is a behaviour of the system answering the question, and a business influences it only by what its pages say. The second is treating it as a trick. Nothing in the site’s sources suggests a way to game the sub-questions, and the measured levers are about clarity, structure and evidence.

The third is assuming every assistant fans out in the same way. The ranking article reports different citation patterns for the AI Overview, for ChatGPT, Gemini and Copilot, and for Perplexity, and this page makes no claim about how any one product implements the step. The fourth is assuming that ranking no longer matters. The figures above say it matters less for the AI Overview than it did, not that it has stopped mattering.

Where it goes wrong

One failure is writing a broad page that touches every angle and answers none of them clearly. Fan-out looks for the clearest source for one narrow point, and a page that is mildly relevant to ten sub-questions is rarely the clearest for any of them. Another is chasing a sub-question that has nothing to do with what the business actually does, which adds a page and no evidence of expertise.

A third is publishing claims that cannot be supported, since a system choosing a source for each sub-question is also choosing which claims to attribute, and a vague or unsourced claim is the easiest one to skip. The last is expecting a click. A citation can be won without one, and a query answered on the results page may send fewer visits even when a page holds its position.

Where it appears in an Audit

Fan-out shows up in an audit as a question about content: for the handful of topics that matter most to the business, which narrow questions sit around them, and does an existing page answer each one clearly enough to be chosen. The ranking and AI Overviews articles describe that as checking a site against a different question from the one SEO used to ask, and they note that most of the work is not a rewrite but an audit of what a site already has.

An audit starts with a discovery call about where a site stands today. If you want your pages checked against generative AI search directly, you can get an audit to begin one. For the wider picture, read how AI citation works, then the two articles this page draws on: how AI Overviews are changing what counts as SEO and why ranking on Google is not enough.