ChatGPT Query Fan-Out: What 15 Million Queries Mean for AEO

17 October 2026 · The Cresia team

ChatGPT rarely answers a question from a single search. It breaks the prompt into several narrower queries, retrieves results for each, and assembles the reply from whatever comes back. Ahrefs has now looked at that behaviour at scale, which means AEO and GEO teams can plan for the queries the model actually runs, not only the words a person typed.

Key takeaways

ChatGPT often turns one prompt into several smaller searches, so your page competes for the sub-questions, not just the prompt a user typed.
Ahrefs analysed 66.6 million logged-out GPT-5.6 Luna responses and 15 million fan-out queries from August and September, the largest public look at this behaviour so far.
Plan content around question clusters: definitions, comparisons, constraints and proof points that a model would search for on its own.
Measure by sampling prompts, recording the sub-queries and citations, and tracking them over time. Single-prompt rank checks mislead.
Treat the findings as a strong directional signal from free, logged-out sessions, not a law that holds for every account, market or model version.

What the Ahrefs fan-out study looked at

Ahrefs Blog reported on a dataset of 66.6 million free, logged-out GPT-5.6 Luna responses collected in August and September, along with the 15 million fan-out queries that sat behind them. Those queries are the searches the model generates for itself before it writes an answer. They are usually not the words the user typed, and they are usually not phrased the way a person would phrase them.

That is the useful part. Until now, most AEO work has been built on guessing. A team would pick fifty prompts, paste them into a chat window, and note whether the brand showed up. The prompt is only the first step in the chain. What decides citation is the set of searches that follow it.

Two cautions apply before anyone rebuilds a content plan around the study. First, the sample is free, logged-out traffic. Paid accounts, logged-in users with memory, and different regions may behave differently. Second, the model version matters. Retrieval behaviour changes when a vendor changes the model, and a pattern from August and September can drift by spring. Read the full write-up at Ahrefs for the exact figures and method, and check your own prompts against it before you commit budget.

Why fan-out changes how you plan AEO and GEO content

The old mental model of search is one query, one ranked list, one click. Fan-out breaks that model. A single prompt such as a request for a project management tool for a 200-person agency can turn into separate searches for pricing, integrations, security certifications, comparisons with two named rivals, and reviews from similar companies. Your page might be a strong answer to one of those and absent from the rest.

That shifts the unit of competition. You are no longer fighting for one head term. You are fighting for a handful of narrow questions, and the model picks sources per question. A brand with one big, vague overview page tends to lose to a brand with several focused pages that each settle one question cleanly.

It also explains something many teams have seen without a name for it: being cited for a prompt you never targeted. If a model asks its own question about data residency and your documentation answers it directly, you can show up in an answer about a topic your marketing never mentioned. The reverse happens too. You can rank well for the obvious prompt and still be left out, because the model's follow-up searches landed on other pages.

The practical consequence is that keyword lists need a second layer. Alongside the terms people search, you need the terms a model searches on their behalf: more literal, more specific, more likely to include a year, a constraint or a comparison.

How AI answer engines choose and cite sources

No vendor publishes its full selection logic, and any claim to know the exact weighting should be treated with suspicion. What can be said is based on how retrieval-based systems work in general, and what the vendors document themselves.

The pipeline has roughly four stages. The model interprets the prompt and decides whether it needs fresh information. If so, it generates one or more search queries. A retrieval layer fetches candidate pages, often from a search index plus its own crawler. Then the model reads passages from those pages and writes an answer, citing some of them.

Three things follow from that shape.

Eligibility comes first. A page that cannot be crawled, renders its content only after heavy client-side scripting, or sits behind a login never gets read. Check the vendor's crawler documentation and your robots rules before anything else.
Passages beat pages. The model reads fragments. A paragraph that states a fact, names its subject and gives its scope can be lifted whole. A paragraph that depends on three earlier paragraphs for meaning cannot.
Citations favour the clearest answer to the sub-question, not necessarily the strongest domain. A smaller site with a precise answer can beat a large site with a general one, though authority still counts when several pages are equally clear.

Search Engine Land's generative engine optimisation coverage and the Conductor benchmarks report are both worth reading for how brands are showing up in practice. Treat them as snapshots, not rules.

What to change on your site

Start with the work that has the clearest payoff and the lowest risk of wasted effort.

Map the sub-questions for your top commercial topics. For each topic, list what a careful buyer would ask in order to decide: price model, limits, integrations, security, who it suits, who it does not, how it compares. Those are your likely fan-out queries.
Give each sub-question a findable home. That can be a section with a descriptive heading on a larger page, or a short standalone page. The heading should read like the question or its direct answer.
Answer first, then explain. Put the plain answer in the first two sentences under the heading. Add the nuance, the exceptions and the evidence after.
State facts in full sentences with the subject named. Write that a plan includes single sign-on, not that it is included. Pronouns and references to above make passages useless on their own.
Publish the unflattering facts. Models search for limits and drawbacks. If your own page does not cover them, a competitor's comparison page or a forum thread will.
Keep specifics current. Prices, supported integrations and certifications go stale. Put a visible updated date on pages where freshness matters, and actually update them.
Fix the plumbing. Server-rendered text, clean canonical tags, working sitemaps and structured data that matches the visible content. None of this is new, but fan-out makes the cost of getting it wrong higher because there are more queries to miss.

A worked example helps. Suppose a payroll software company wants to be named for the prompt best payroll software for a 300-person manufacturer. Likely sub-queries include shift-differential handling, multi-state tax compliance, integration with common timekeeping systems, and implementation time. The company's current site has a single product page and a gated PDF. The fix is not more copy on the product page. It is four short, direct pages or sections, each answering one of those questions with specifics a buyer can verify.

How to measure fan-out visibility

Rank trackers were built for a single query and a single list. They do not capture what happens when one prompt becomes five searches and the answer cites three sources. You need a sampling approach instead.

Build a fixed prompt set of 50 to 150 prompts that reflect real buyer questions, grouped by topic and stage. Run them on a schedule, record the answer text, the cited sources and, where the interface shows it, the sub-queries the model ran. Keep the set stable so changes mean something. Add new prompts rather than swapping old ones.

What to track	How to capture it	Why it matters	Likely owner
Citation rate per topic	Share of sampled answers that cite your domain	Shows whether you are eligible and relevant at all	SEO or AEO lead
Mention without citation	Brand named in text but no link	Signals awareness without a retrievable source page	Content lead
Sub-query coverage	Which fan-out queries return your pages	Reveals the specific questions you fail to answer	SEO and content
Competitor co-citation	Which rivals appear alongside you	Shows who you are being compared with	Product marketing
Referral traffic from AI tools	Analytics segments by referrer	Ties visibility to sessions and conversions	Analytics lead
Answer accuracy	Manual check of claims about your brand	Catches wrong prices, features or availability	Brand or PR

Two notes on method. Answers vary run to run, so sample each prompt several times and look at the pattern, not one result. And keep referral tagging tidy: if AI traffic lands in a general referral bucket, you cannot connect it to outcomes. A written tracking specification that names the AI referrers and how they are classified keeps this consistent across teams. Teams that want to watch AI search visibility alongside their wider media reporting can look at MediaPilot.

What to ignore and what not to do

Most bad AEO advice shares one flaw: it treats a model's behaviour as a fixed rulebook. Here is what to leave alone.

Do not chase every conceivable sub-query. The number of possible fan-out searches is huge. Cover the ones that decide a purchase and stop.
Do not stuff pages with the awkward phrasings models generate. Write for the question, in normal language. Pages that read like a keyword list lose the human reader and gain nothing reliable.
Do not build thin pages at scale. Hundreds of near-identical question pages are easy to produce and easy for search systems to discount.
Do not rewrite everything after one study. One dataset on one model in one period is a reason to test, not a reason to restructure the site.
Do not rely on a tool that reports a single visibility score without letting you see the prompts and the raw answers behind it.
Do not hide key facts in images or PDFs that retrieval systems may not read well.

The long-standing basics still apply. Pages that are useful, accurate and clearly written serve both a person and a retrieval system. Fan-out raises the reward for covering a topic fully and plainly. It does not replace quality.

Who does what: a 30-day plan

Fan-out work crosses team lines, so assign it explicitly. SEO owns the prompt set and the sub-question map. Content owns the rewrites and new sections. Analytics owns referral classification and reporting. Product marketing owns the comparison and limitation pages, because those need honest input from people who know the product. Teams organised around growth work often find the coordination harder than the writing.

In week one, build the prompt set and run a baseline. In week two, map sub-questions for your three most valuable topics and compare against existing pages. In weeks three and four, ship fixes for the largest gaps and re-run the same prompts. Then review: did citations for those topics move, and did any wrong claims about your brand appear? Expect slow and uneven change, and expect some of it to be noise.

Frequently asked questions

What is query fan-out in ChatGPT?

It is the step where the model turns one prompt into several narrower search queries before it writes an answer. Each query retrieves its own set of pages, and the final reply draws on the combined results. Ahrefs Blog studied 15 million of these queries to see how they look in practice.

Does fan-out mean keyword research is dead?

No. People still search with keywords, and Google still drives most organic discovery for many sites. Fan-out adds a second layer: the questions a model asks on a user's behalf. Use both, and use your existing search data as a starting point for guessing those sub-questions.

Can I see which fan-out queries ChatGPT ran for my brand?

Sometimes. Some interfaces show the searches the model performed, and some third-party tools sample them. Coverage varies and changes with product updates, so check the vendor's own documentation or interface for what is currently exposed. Where you cannot see them, infer them from the questions a buyer would ask.

Do the Ahrefs findings apply to Gemini, Perplexity and Google AI Overviews?

Not automatically. The study covers one model's free, logged-out responses. Other systems retrieve differently and may fan out more or less. The sensible approach is to run your own prompt set across the engines your buyers use and compare.

How long before changes show up in AI answers?

It depends on how often the relevant index and crawler revisit your pages, and on how the model weighs fresh material. Some changes appear within days, others take weeks, and some never register. Re-run the same prompts on a regular schedule and judge the trend over a quarter rather than a week.

Sources

https://ahrefs.com/blog/query-fan-out-chatgpt/
https://searchengineland.com/topic/generative-engine-optimization
https://www.conductor.com/academy/aeo-geo-benchmarks-report/