How Query Fan-Out Works: Worked Example and Coverage Matrix

Query fan-out is how AI search answers a prompt: the engine splits it into narrower searches, reads results for each, and writes one answer from them. Google documents it for AI Mode and AI Overviews, and ChatGPT and Perplexity search in comparable ways. Pages get retrieved and cited against those sub-queries, so plan content from them.

A keyword list assumes a searcher types one query and picks from ten links. A fan-out means the engine picks for them, across several searches you never saw. The practical change for SEO teams is where you look for gaps: in the searches behind the answer, where your pages may not rank at all.

One prompt, ten searches

One buyer question can turn into ten or more searches, each covering a condition hidden in the prompt. The example below uses a fictional prompt about payroll software.

Prompt: "What's the best payroll software for a 40-person restaurant group with tipped staff?"

The sub-queries below are modeled. We wrote them from Google's description of fan-out ("breaking down your question into subtopics") and the query types the Citlyze fan-out simulator uses: reformulations, implicit questions, comparisons, follow-ups and related entities. No engine published these exact searches for this prompt.

#	Modeled sub-query	Type	Content that would answer it
1	best payroll software for restaurants 2026	Reformulation	Category roundup, review sites
2	payroll software tip pooling and tip credit	Implicit requirement	Feature docs, help center
3	restaurant payroll multiple locations	Implicit requirement	Product page for multi-entity payroll
4	payroll software pricing per employee	Implicit requirement	Pricing page
5	payroll that integrates with restaurant POS	Implicit requirement	Integrations directory
6	Gusto vs ADP for restaurants	Comparison	Comparison pages, reviews
7	restaurant payroll software reviews reddit	Social proof	Forums, review platforms
8	FLSA tip credit rules for payroll	Compliance follow-up	Government and legal explainers
9	how to switch payroll providers mid-year	Follow-up	Migration guides
10	payroll for hourly shift workers overtime	Related need	Feature docs, blog posts

A vendor that ranks first for "restaurant payroll software" might still miss half of this list: no public pricing page, a tip-pooling explainer buried in a PDF, no integration page for the POS systems restaurants use. Each gap is a sub-query where another site supplies the answer, and often the citation.

How AI Mode, ChatGPT and Perplexity differ

Each engine documents a different amount, and each exposes a different amount of its searching. Google describes parallel searches; OpenAI describes searches a reasoning model can repeat; Perplexity reports counts at most.

Engine	What the operator documents	What you can observe
Google AI Mode	AI Mode "uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously" (Google). Deep Search "can issue hundreds of searches".	No query data in Search Console. The Gemini API returns the searches it ran in a webSearchQueries field (Gemini API reference), which is the closest public proxy
Google AI Overviews	"Both AI Overviews and AI Mode may use a 'query fan-out' technique" (Google Search Central)	Impressions by page only
ChatGPT	Reasoning models can "perform web searches as part of its chain of thought, analyze results, and decide whether to keep searching" (OpenAI)	Through the API, a web_search_call item that "will usually (but not always)" include the queries. No documented export from the consumer app
Perplexity	Deep Research reports how many searches it ran, 21 in the documented example (Perplexity)	Search results and a count. The query strings are not returned

The difference in shape matters for planning. A planned, parallel fan-out rewards pages that match the obvious sub-topics of a question. An iterative search that decides "whether to keep searching" rewards pages that answer the follow-up the first result raised, such as the migration question in row 9.

Google's own help documentation uses the same language. Its guide to impressions and clicks says AI Mode "groups the user's question into subtopics and searches for each one simultaneously". How many sub-queries one prompt produces varies by engine, prompt and mode: Google says Deep Search can issue hundreds, Perplexity's documented example ran 21, and none of the engines publishes a typical count for a standard answer.

Why you can rank first on Google and still not get cited

You rank for the query a person types, while the answer is assembled from results for sub-queries you may not rank for, and from passages the engine judged relevant to each one.

Three mechanisms stack up:

• The prompt is not the query. A long question produces searches that look nothing like your head term. Row 8 above sends the engine to compliance sources that a payroll vendor's site may not rank for at all.
• Selection happens per passage. A page can rank on a topic and still lose the passage-level comparison for one condition in the question, which the reranker test shows in a controlled example.
• Engines search different parts of the web. In our August answer study of 200 buyer prompts, ChatGPT and Perplexity shared no cited domain on 77% of prompts.

Google is also putting AI Mode in front of more searchers. Since late August, some AI Overviews expand into a full AI Mode response with the "Ask anything" box already open. A Google spokesperson said, as reported by Search Engine Roundtable, that "for some queries, AI Overviews may dynamically expand for topics where our systems determine it's most useful for people." More expanded answers invite more follow-up questions, and Google counts each follow-up as a new query, with its own fan-out.

Google's eligibility rule still applies underneath all of this. To appear as a supporting link in AI Overviews or AI Mode, "a page must be indexed and eligible to be shown in Google Search with a snippet", and Google says no special optimization beyond that is needed.

Building a fan-out coverage matrix

List the sub-queries behind your priority prompts, check which of your pages answers each one and where it ranks, and work the gaps in order of how often each sub-query recurs.

1. Choose 10 to 20 prompts your buyers ask. Finding buyer prompts covers the research.
2. Collect sub-queries for each prompt. Use observed queries where an engine exposes them (the Gemini and OpenAI APIs, or a tracker that captures them) and modeled ones where it does not. Label every row as observed or modeled.
3. Merge near-duplicates into clusters, and count how many of your prompts each cluster appears under.
4. For each cluster, record the URL on your site that answers it, your Google position, the top three ranking domains, and the domains AI answers cite for the parent prompt.
5. Classify each row: answered and ranking, answered but not ranking, or not answered.
6. Rank the gaps by recurrence across prompts, then by the commercial value of the parent prompts.
7. Fix the top gaps by improving an existing page first. Add a new page only when no current page can carry the answer.

Sub-query cluster	Observed or modeled	Your page	Google position	Cited for the parent prompt	Gap	Action
Tip pooling and tip credit	Modeled	Help article	Not in top 100	Review site, government page	Answered, not ranking	Move the answer onto the product page and link the help article
Pricing per employee	Modeled	None	None	Two competitors' pricing pages	Not answered	Publish pricing, or a clear pricing explainer
POS integrations	Observed (API)	Integrations page	4	Your page	Answered and ranking	Keep current; add the missing POS names

The rows above are placeholders for the fictional vendor, to show the format. Your own matrix needs your own rankings and citations.

Step 7 is the easy one to get wrong. A fan-out list looks like a publishing plan, and producing one thin page per sub-query drifts toward what Google's spam policies call scaled content abuse. Answer the conditions of a question inside the pages you already have, and add a page only where no existing page fits. The semantic SEO approach, one strong page per entity and job with the conditions answered inside it, fits fan-out better.

To see a modeled fan-out for your own prompt, paste it into the free query fan-out simulator. It returns 8 to 14 sub-queries grouped by type, each labeled with the kind of content likely to answer it.

Tracking fan-out over time

Capture sub-queries on a schedule, keep observed and modeled data in separate columns, and watch which sub-queries recur, because single captures vary.

• Fan-out changes by run, by user context and by date, so one capture is an anecdote. Count how often a sub-query appears across weeks before acting on it.
• Track your Google rank for the most frequent sub-queries weekly. Those are the searches the engine runs before it answers.
• Measure the outcome on the parent prompt: mention and citation rate in the answers, alongside rankings.
• Note engine changes. When AI Overviews started expanding into AI Mode, the number of follow-up searches per session could change without any change on your site.

Search Console will not help with the sub-queries themselves. Its Generative AI performance report shows impressions by page, country, device and date, with no query data.

What is observed, and what is inferred

Google, OpenAI and Perplexity confirm that their engines can issue several searches for one question. The exact searches behind a consumer answer, and the link between ranking for them and being cited, are inferred.

Claim	Status	Basis
AI Mode and AI Overviews fan out a question into several searches	Documented	Google's AI Mode announcement and Search Central documentation
The exact sub-queries AI Mode ran for your prompt	Not observable	Search Console reports impressions by page, without queries
The searches Gemini ran through the API	Observed, API only	webSearchQueries in grounding metadata
The searches ChatGPT ran through the API	Observed most of the time, API only	OpenAI says queries are "usually (but not always)" included
Perplexity's search strings	Not observable	Counts only, for Deep Research
Ranking for a sub-query raises your odds of a citation	Inferred	Consistent with how retrieval works; no engine publishes the weighting
The sub-queries in this article's example	Modeled	Written for illustration, not captured

Keep that last column in every report you share. A modeled sub-query presented as observed data is how teams end up rewriting pages for searches no engine runs.

Map your three most valuable prompts first

Model their fan-out and check which sub-queries you cannot answer today. The free query fan-out simulator does the first step in a minute. Citlyze's query fan-out tracking then captures the sub-queries engines report for your own tracked prompts on every plan, keeps modeled variants labeled apart from observed ones, and on Growth and Pro runs a weekly Google rank check on the top 25 or 50 fan-out queries. Plans start at $29 a month with a 14-day trial, no card.