Query fan-out: how AI turns one question into many searches

When you ask an AI engine a question, it may not search for that exact phrase. It can break your question into several smaller sub-queries - a process Google calls query fan-out - search for each one separately, and synthesize the results into one answer. The practical takeaway: to get cited, your content has to match the sub-questions the engine generates, not just the headline query a person typed.

What is query fan-out?

Query fan-out is the set of related queries an AI model generates to gather enough information to answer you. Google says AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches. A question like "how do I fix a lawn full of weeds" might fan out into "best herbicides for lawns," "remove weeds without chemicals," and "how to prevent weeds in lawn," then pull sources for each and weave them together.

So the single question a buyer types can become several searches under the hood. Each sub-query has its own set of candidate sources, and your brand can be evaluated separately for each one. You can win the answer by being the best source for several sub-queries - even if no single page targets the original phrase.

Curious whether AI engines name you when buyers ask about your category? Get a free visibility audit and see where you stand across ChatGPT, Gemini, Perplexity and Google AI Overviews.

Why this changes how you should write

Optimizing for one keyword phrase made sense when one query returned one ranked list. With fan-out, a buyer's question fragments into intents you didn't literally write for. A page that comprehensively covers a topic - definitions, comparisons, costs, objections, edge cases - matches more of the fan-out and tends to be pulled into more answers. A thin page built around a single keyword matches one sub-query at best.

This can also be why a competitor with no page targeting your exact term can still beat you in the answer: their content happened to answer three of the sub-queries cleanly, and yours answered one.

How to find the sub-queries

1. Start from the buyer's real question, not a keyword - e.g. "what's the best AEO tool for a mid-market brand?"
2. Decompose it the way a model would: "what is an AEO tool," "AEO tools for mid-market," "AEO vs GEO tools," "how much do AEO tools cost," "AEO tool alternatives to [competitor]."
3. Check coverage: for each sub-query, do you have a clear, self-contained answer somewhere in your content? Gaps are where competitors win.

How to structure content to win the fan-out

• Go comprehensive, not narrow. Cover the cluster of related sub-questions on a topic rather than chasing one phrase.
• Use question-shaped headings. H2s phrased as the sub-questions make it obvious - to both readers and models - that you answer them.
• Answer each sub-question in its own self-contained block a model can lift without needing the rest of the page.
• Interlink related pieces so the engine can assemble a fuller picture of your coverage.

Fan-out is one reason strong rankings don't guarantee citations - the engine may not be matching your ranked page to the typed query so much as matching many sources to many sub-queries. We unpack that gap in strong SEO, invisible in AI search.

"Citations can be awarded per sub-query, not per page - so being the clearest source on one fragment can pull you into an answer you never targeted."

A worked example: one query, twelve sub-queries

Abstract advice about fan-out only clicks once you watch a single question shatter. Take a buyer typing "is a CDP worth it for a 50-person ecommerce brand?" into an AI engine. That is one query. Here is a realistic set of sub-queries a model might fan out into before it writes a word:

• what is a customer data platform
• how much does a CDP cost per month
• CDP vs data warehouse - what's the difference
• do small businesses need a CDP
• CDP pricing for ecommerce
• best CDP for small ecommerce brands
• CDP implementation time and effort
• what data sources does a CDP connect to
• CDP alternatives for small teams
• CDP ROI examples ecommerce
• can you build a CDP in-house instead of buying
• signs you've outgrown spreadsheets and need a CDP

Notice that the buyer's literal phrase - "worth it for a 50-person brand" - appears in none of them. The engine translated intent (should I buy this, at my size, for my money) into several answerable fragments. If your page only argues "CDPs are worth it," you match maybe two of those twelve. A rival who documented pricing bands, an in-house-versus-buy breakdown, and an implementation timeline matches seven or eight and has far more chances to be woven into the answer. You are not competing for one ranking anymore. You are competing twelve times in parallel, and your score is how many of those rounds you win cleanly.

Map the sub-question tree before you draft

Because one query fans into many, the unit of planning is no longer the keyword - it is the topic's sub-question tree. Before writing, spend twenty minutes building that tree so you know what you are on the hook to answer. A repeatable way to do it:

1. Write the buyer's real question in their words. Not "CDP pricing" but "is a CDP worth it for a 50-person ecommerce brand?" The messier and more human, the better - that is what people actually ask engines.
2. Decompose along five intents worth covering: definition ("what is X"), comparison ("X vs Y"), cost ("how much does X cost"), objection ("do I actually need X"), and edge case ("X for [my specific situation]"). Each intent tends to spawn its own sub-query.
3. Ask the engines directly. Pose your headline question to ChatGPT, Perplexity, and Google AI Overviews, then read which facets each answer covers and which sources it cites. That is a rough readout of the facets the engine covers, and it costs nothing.
4. Score your current coverage. For each sub-question, mark whether you have a clear, self-contained answer somewhere on your site. Green, thin, or missing. The thin and missing rows are where a competitor is most likely being cited instead of you.

The output is a checklist, not a vibe. You should finish knowing you are covering, say, nine of twelve sub-questions and that the three gaps are pricing bands, in-house-versus-buy, and implementation time. That is a content brief you can actually assign.

Depth belongs on one page, not scattered across ten

A tempting misread of fan-out is "publish a separate thin post for every sub-query." That can backfire. Ten shallow pages each answer one fragment and force the engine to stitch across your whole site to assemble a picture - which it may not bother to do. A single comprehensive page that answers the full cluster gives the model one confident, coherent source to lift from, and it matches far more of the fan-out at once.

Structure that page so any block can stand alone. Concretely:

• Question-shaped H2s and H3s. "How much does a CDP cost?" beats "Pricing." The heading tells both the reader and the model exactly which sub-query the section resolves.
• Self-contained answers. Each section should make sense if it were the only paragraph a model quoted. Restate the subject instead of relying on "it" or "as mentioned above." Fan-out pulls fragments out of context, so write fragments that survive being pulled.
• Answer-first, then detail. Lead the section with the direct answer in the first sentence, then justify it. Passages that resolve the sub-query up top tend to be easier for a model to lift.
• Reserve separate pages for genuinely separate topics, and interlink them so the engine can see the fuller cluster. Depth clusters on a page; breadth clusters across linked pages.

How fan-out decides which citations you earn

Fan-out is one likely mechanism behind a frustration a lot of teams hit: they rank well, yet the AI answer cites someone else. One reason can be that the engine never matched your ranked page to the typed query. It may run several sub-queries and can pick, for each one, a source that answers that fragment cleanly. In effect, citations can be awarded per sub-query rather than per page.

Three consequences worth internalizing:

• Partial wins are real wins. You do not need to own the whole answer. Being the clearest source on "CDP implementation time" can earn you a citation inside an answer whose headline question you never targeted. Cover more sub-questions well and you appear in more answers.
• The answer can be a stitched multi-source quilt. A single AI response can cite three or four different sites, each supplying the fragment it answered best. Your goal is to be the best answer for as many of those fragments as you can, not to shut everyone else out.
• A cited passage has to be liftable. If your pricing answer is buried in a paragraph that also compares two products and tells a customer story, the engine has nothing tidy to quote for the "how much does it cost" sub-query. Clean, scoped, factual blocks are what tend to get pulled and attributed.

Practically, this is why monitoring belongs at the sub-query level. Tracking whether you rank for one phrase tells you little. Tracking which sub-questions engines are citing you for - and which ones send them to a competitor - tells you exactly which block to write next.

How ChatGPT's fan-out changed in 2026: the site: operator

Fan-out is not only a Google behavior. In mid-2026, two independent studies reported ChatGPT running many more sub-queries restricted to a named website with the site: operator, so that some of its searches look inside chosen sites rather than across the open web.

• Writesonic (study published July 16, 2026) ran 50 prompts across 16 categories once per configuration, reading ChatGPT's backend responses. Queries using site: made up 12.6% of searches for GPT-5.5 at Medium effort and 59.3% for GPT-5.6 Sol at Medium, and 10.7% against 71.2% at High. The authors describe the result as a map rather than a precise measurement, since it reflects a single user, a single run and a single point in time.
• Promptwatch reported that on August 8, 2026, queries using site: rose from 0.37% to 16.8% of ChatGPT's fan-out queries, and that the average number of fan-out queries per response rose from about 1.08 to 1.83.

The two studies measure different things, a small prompt set on specific models in one case and an ongoing tracking dataset in the other, so their numbers are not directly comparable and neither is a ChatGPT-wide rate. OpenAI has not confirmed the change or explained how ChatGPT chooses which sites to query.

What it can mean for fan-out planning: covering the sub-questions is still the job, but where each answer lives may matter more. If ChatGPT picks sites first and then searches inside them, it can help to hold a clear answer to each sub-question on your own site, and to know which third-party sites in your category its sub-queries name. In Writesonic's test, the share of citations going to brands' own sites at Medium effort rose from 47.2% under GPT-5.5 to 58.1% under GPT-5.6 Sol.