Query fan-out: how AI search picks its sources (SEO, AEO and GEO guide)

TL;DR: When you ask Google AI Mode, AI Overviews, ChatGPT or Perplexity a question, the engine does not search for your words. It splits the prompt into 5 to 10 smaller searches, pulls passages for each one, and writes an answer from what it found. That split is called query fan-out. If your pages answer the smaller searches, you get cited. If they only target the headline keyword, you do not. This guide explains how fan-out works, how to measure it, and how to write pages that survive it.

What is query fan-out?

Query fan-out is the step where an AI search engine turns one user prompt into several background searches, runs them against a web index, and combines the results into a single answer with citations.

Ask an AI engine "what is SEO?" and it will rarely search those three words. It guesses what you actually want to know and runs searches like:

• seo definition
• benefits of seo for a small business
• on-page vs off-page seo
• technical seo basics

Each of those searches returns a handful of passages. The model reads them, picks the ones it trusts, and writes the answer. The pages it read become the citations under the answer.

Why engines do this

1. To avoid making things up. A language model on its own answers from memory. Fetching live passages gives it something to quote, and the citation is the proof.
2. To get past the training cutoff. Prices, algorithm updates, and news change weekly. Fan-out lets the model read today's web.
3. To save the user work. A person would run five searches over ten minutes. The engine runs them in parallel in two seconds.

SEO, AEO and GEO: what changes

Fan-out is why the three disciplines now sit side by side. The same page has to work for a classic results page, for a direct answer box, and for a generated answer that cites sources.

	SEO	AEO (answer engines)	GEO (generative engines)
Goal	Rank in the top ten for a keyword	Be the direct answer: featured snippet, voice result, quick answer	Be named or cited inside an AI-written answer
Unit optimised	The page, its links, its keywords	The paragraph that answers the question, plus schema	The passage (chunk), the entity, and what third parties say about it
What you measure	Clicks and impressions	Share of answer boxes, zero-click impressions	How often you are cited across repeated runs, entity mention share
What triggers it	Index and ranking algorithm	Knowledge graph and entity extraction	Fan-out subqueries and passage retrieval

Naive fan-out vs real fan-out

A lot of tools "simulate" fan-out by asking a model once: "what would you search for if someone asked X?" and printing the list. That is a guess from memory, not a measurement. It fails in two ways.

1. Asking a model is not the same as grounding

The list you get back comes from the model's training data. In production, engines decide whether to search at all based on the prompt (is it time-sensitive? is it something the model already knows?). A one-shot prompt cannot reproduce that decision.

2. Later searches depend on earlier results

Real fan-out is a chain. If the first search for "Search Console update 2026" returns a passage about a new branded queries filter, the next search becomes "Search Console branded query filter documentation". A single static list cannot predict that turn because it never saw the passage.

Why you need several runs

Generation is probabilistic. Run the same prompt once and a subquery may appear by chance. Run it eight times and you can see which subqueries show up every time and which were noise. Eight parallel runs is the practical minimum for numbers you can act on.

Three numbers that matter

Subquery probability

The share of runs in which a subquery (or a cluster of near-identical subqueries) appeared.

probability = runs where it appeared / total runs x 100

Rule of thumb: under 20% (one run in eight) is noise. Spend your content budget on subqueries above 60%.

Average position

Where in the sequence the engine ran the subquery. Early subqueries set the context for everything after them, so a subquery that always runs first has more influence than one that runs fifth.

RRF score

Reciprocal rank fusion folds frequency and position into one number so you can sort a list of subqueries by influence:

RRF = sum over runs of 1 / (k + position), with k = 60

A subquery that appears in every run near the top gets a high score. One that appears twice near the bottom does not.

How the main engines fan out

Google AI Overviews

• Triggers on ordinary searches, mostly informational and high-intent ones.
• Small fan-out: usually 2 to 5 subqueries.
• Stays in the language and region of the original search.

Google AI Mode

• Built for long, exploratory prompts and follow-up questions.
• Large fan-out: 5 to 10 or more subqueries per run.
• Will pull sources in other languages when it needs them.

ChatGPT Search

• Searches selectively. If the model thinks it already knows, it answers from memory. Fresh or time-bound prompts trigger a search.
• Subqueries tend to be longer and more specific, to compensate for a smaller index.

Writing pages that survive fan-out

Engines do not read your page as one document. They cut it into passages, embed each one, and match passages to subqueries. So the unit you are optimising is the passage under each heading.

1. Make every section stand alone. The text under each H2 or H3 should make sense if it is lifted out on its own. No "as we said above".
2. Answer first, explain second. Put a 40 to 60 word direct answer right under the heading, then the detail.
3. Add schema. Article, FAQPage, Product and Organization JSON-LD tell the engine what the entities on the page are and how they relate.
4. Work on what others say about you. For "best X" and comparison prompts, engines look for agreement across several third-party pages. Your own page cannot win those subqueries alone; reviews, listicles and press on other domains can.

A four-step workflow

Step 1: pick seed prompts

Start with broad prompts a real customer would type, not keyword strings. "What are the best SEO platforms for an e-commerce store?" beats "best seo platform ecommerce".

Step 2: run the fan-out eight times

Run each seed prompt through at least eight parallel runs. Drop subqueries that appear once. Sort the rest by RRF score.

Step 3: sort subqueries into two buckets

• Owned content. Definitions, how-tos, feature breakdowns. Write a section on your own site for each one.
• Third-party consensus. "Best of" lists, reviews, neutral comparisons. Find the external sites that keep appearing as sources for those subqueries and get listed there.

Step 4: publish, then re-run

Ship the new sections with clear headings, direct answer blocks and tables where the data is tabular. Run the same eight-run audit again and compare citation share before and after.

Where RankNexus fits

RankNexus tracks AI visibility across ChatGPT, Gemini, Perplexity and Google AI Overviews for your domain, shows which prompts cite you and which cite competitors, and turns the gaps into content briefs and technical fixes. The fan-out method above is the reasoning behind that: rank for the hidden searches, and the visible answer follows.

Start with a free audit of your site to see where you stand today.