Skip to content

Guides

Reading fan-out data: what AI sub-queries say about buyer intent

Gemini 3 averages 10.7 sub-queries per prompt and 26.4% of them name a brand, so read a month of fan-out lists by share and wording rather than by volume.

Five flat discs of varying sizes converge toward a single point via thin lines, one disc highlighted in blue

SEO consultant, and more recently GEO, at MentionBox

A fan-out list is buyer-intent evidence, because every sub-query in it is a condition the engine decided to check before naming anyone. Seer Interactive's test of 501 prompts on the Gemini 3 API, with grounding forced on, returned an average of 10.7 sub-queries per prompt, from a low of 3 to a high of 28, and 26.4% of them already carried a brand name. The reading that pays is not one list but the shape of a month of them: which conditions keep coming back, which ones name a competitor, and which ones no page of yours could answer.

This page is for anyone who already collects fan-out queries, through a model API or a visibility API, and now has a spreadsheet of machine-written sub-questions that nobody has turned into a decision. It covers what a fan-out list records, the wording signals that carry intent, the counts worth computing, what the data cannot tell you, and how a sub-query becomes a page. For the mechanism itself and the list of APIs that return it, see our explainer on query fan-out.

What is in a fan-out list

The strings come from the model, not from a search box, and their shape shows it. Seer measured Gemini 3 fan-outs at 6.7 words on average, between 2 and 17, with 95% of them carrying no global search volume. DataForSEO's analysis of 100,000 ChatGPT prompts and 100,249 fan-out queries, published on May 12, 2026, puts the same gap in characters: user prompts cluster between 31 and 60 characters, the fan-out queries they produce between 61 and 90, and a fan-out query shares 54% of its words with the prompt behind it. The engine keeps about half of your buyer's wording and writes the other half itself. That written half is the intent data.

How often a prompt fans out at all depends on the engine and on what was counted. DataForSEO found a fan-out behind 47.5% of the ChatGPT prompts in its sample. Ahrefs, looking at 66.6 million free logged-out ChatGPT responses and their 15 million fan-out queries from August and September 2026, found fan-out queries in 14.4% of responses, averaging about 1.6 each. The two figures measure different populations, one a tracked prompt set and the other everything people send to a chatbot, most of which never needed a search.

Volume per engine sets how much evidence a single prompt can give you. Surfer's research puts the averages at 10.7 sub-queries on Gemini 3, 6.8 on Grok, 2.1 on ChatGPT and 1.4 on Perplexity. A Perplexity prompt hands you one sub-question to read; the same prompt on Gemini hands you ten, which is why a fan-out programme usually starts on the Google side even when ChatGPT is the engine the business cares about.

How the wording of a sub-query carries intent

Six signals in the string do most of the work. Each one has a published frequency, so you can tell a pattern in your own data from a quirk of one prompt.

Signal in the sub-queryPublished frequencyWhat it says the engine is doing
A year, such as "2026"22.9% of ChatGPT fan-outs in September 2026, up from 17.6% in August (Ahrefs); 21.3% of Gemini 3 fan-outs carried a year (Seer)filtering for recency, which drops undated pages out of the candidate set
"price", "cost", "pricing"8.9% of ChatGPT fan-outs in September 2026, against 6.0% in August, a relative rise of 47% in one month (Ahrefs)costing an option, so the figure has to sit on a page a crawler can read
"official"13.1% in September 2026, from 10.1% in August (Ahrefs)going to whoever owns the fact rather than to a roundup about it
The site: operator6.8% of ChatGPT fan-outs, with 75.9% of the most common site: searches pointed at a brand's own site (Ahrefs)the entity is already chosen and its own pages are being read
A brand name anywhere in the string26.4% of Gemini 3 fan-outs (Seer)checking a shortlist instead of building one
"best", "vs", "alternatives""best" in 5.3% of ChatGPT fan-outs in September 2026 (Ahrefs)still at category level, with no shortlist formed

Frequencies from Ahrefs' study of 66.6 million ChatGPT responses from August and September 2026 and from Seer Interactive's test of 501 Gemini 3 prompts, run on November 21, 2025.

Read down that list and the funnel position falls out of the wording. Category-level strings ("best", "vs", a use case with no vendor in it) mean the shortlist is still open. Entity-level strings (a brand name, site:, "official") mean the shortlist exists and the engine is now verifying claims, usually on the vendor's own pages. Attribute strings sit in between and tell you which specification decided the comparison, which is the single most actionable thing a fan-out list produces: if the engine keeps searching for a limit, a format or a price your pages never state, you lose the sub-query before any ranking happens.

Ahrefs' month-on-month movement is worth watching more than any absolute share. Price searches rose by nearly half in relative terms in four weeks while definition wording fell by about a fifth, which describes a population moving from learning to buying faster than most content calendars move.

The counts worth computing on your own lists

Three numbers turn a dump of strings into something you can put in front of a product team. All three need a period and a fixed prompt set behind them, the way our guide to building a prompt set for GEO sets them up.

  • Cluster share. Sub-queries in a theme divided by all sub-queries collected in the period. Themes, not strings: price, integrations, compliance, support, one competitor name. A theme that holds 20% of the fan-out of your category prompts is where a fifth of the retrieval budget goes.
  • Brand presence rate. Sub-queries naming any brand divided by the total, then split between your name and each competitor's. Against Seer's 26.4% benchmark on Gemini 3, a category where brand-named sub-queries run far below that is one where the engine is still exploring, and a category above it is one where the shortlist is hard to break into.
  • Repetition across runs. How many distinct runs a cluster appears in, out of the runs collected. This is the one that separates a signal from noise, and it is the one most teams skip.

In an analysis we ran in October 2026, a vendor named inside one of these sub-queries was cited for that question 77.1% of the time, against 15.2% when no sub-query named it.

How stable is a fan-out list between runs?

Not stable at the level of the string. Surfer's stability test over 1,600 grounded Gemini runs found only about 27% of fan-out queries staying consistent between runs of the same prompt, with 66% of them appearing exactly once in the whole dataset. Seer saw the same thing from another angle, with only 1% overlap across its full fan-out dataset.

The clusters survive what the strings do not. Surfer found about 90% of fan-outs falling into four or fewer topic clusters per query, which is the practical unit of analysis: a prompt has roughly four things the engine wants to know, and it phrases them differently each time it asks. Our explainer on why the same prompt gives different answers covers the same instability on the answer side.

What fan-out data cannot tell you

Four limits, each with a published reason:

  1. It is not search volume. Seer measured 95% of Gemini 3 fan-out queries at zero global monthly volume. A keyword tool will report nothing for almost every line of your list, and that is the expected result rather than a data problem.
  2. It is not what a logged-in buyer's engine ran. Google's patent Search with stateful chat, published on August 29, 2024, describes generating "alternative query suggestions, supplemental queries, rewritten versions of the user's query, and/or 'drill down' queries", built from the query together with contextual information associated with the user or the client device. A fan-out collected through an API with no history is a neutral baseline, which is what makes it comparable month to month, and also what makes it different from any one person's session.
  3. It is not available on every surface. The model APIs report their searches and the scraping APIs pass them through, but Google's answer surfaces publish no field for them. Anything presented as the fan-out of AI Overviews is reconstructed.
  4. It is not a statement of demand. The engine searched because its own answer needed a fact, not because buyers asked for it. The demand question stays with your sales calls, your support tickets and your search console data.

From a sub-query to a page you write

The mapping is mechanical once the clusters exist. Each family of sub-query is answerable by a different kind of page, and the mismatch between the two is usually what a fan-out audit finds.

Family of sub-queryWhat the engine is checkingWhat answers it
Attribute and limit ("how many X does Y allow")one specification that decides the comparisona specification table on the product page, with the number in text
Price ("Y pricing", "cost per X")a figure it can quotea public price page, with units stated
Comparison ("Y vs Z", "best Y for a small team")who fits a named situationa page built around the situation rather than around your product
Trust ("Y reviews", "Y complaints", "is Y legit")third-party confirmationpages you do not own, which is why review presence is part of the work
Constraint ("Y for regulated industries")an eligibility conditiona page that states the condition and the answer in the same sentence
Entity-owned (site:yourdomain.com, "official Y")your own claim, from youa page on your domain that states the fact plainly, kept current

The two rows most often missing from a content plan are trust and entity-owned. Ahrefs' finding that 75.9% of the most common site: fan-outs pointed at a brand's own site means the engine routinely goes to the source for facts, and a brand whose own pages omit a figure hands that sub-query to whoever stated it.

Where to pull the data

Three routes, in increasing order of what they compute for you.

  • The model APIs, for a neutral baseline at search prices. The Claude API returns one server_tool_use block per search, and Anthropic's web search tool guide states that "simple factual queries typically use 1-3 searches; comparative or multientity research can use 10 or more". You get strings, and the clustering is yours to write.
  • DataForSEO, for volume at the lowest unit price. Its update of December 15, 2025 added fan_out_queries to the LLM Responses API for ChatGPT, Gemini, Claude and Perplexity, and lets the LLM Mentions API search for brand mentions with fan_out_queries as the search scope, which is the closest thing to a query on somebody else's fan-out corpus. Answers cost $1.20 per 1,000 on its standard queue, checked on October 23, 2026, with the counting left to you.
  • API for GEO, for the counts ready-made. Its Fan-Out API returns each sub-question with occurrences, share and a brandPresent flag through POST /v1/fan-outs/top and /by-query/{queryId}, with totalFanOuts, uniqueFanOuts and avgPerResponse as overview totals, and states that reading the ranked list is not metered because fan-outs are captured while prompts run. Answers cost $13 per 1,000 at its first tier, down to $7 at its largest, checked on October 23, 2026, which is about eleven times DataForSEO's rate for the same collection. MentionBox, which publishes this site, has a commercial link with API for GEO.

Eighteen of the 55 APIs we track return fan-out queries, twelve of them at a published price per answer; the full table sits in our explainer on query fan-out, and our ranking scores all 55 on price, engines and data returned.

Four mistakes when reading fan-out as intent

  1. Counting strings instead of clusters. With 66% of sub-queries appearing once, a frequency table of raw strings is mostly a list of hapaxes. Cluster first, count second.
  2. Pulling keyword volume for the list. At 95% zero volume, the exercise returns an empty column and pushes teams back to a keyword plan the engine is not using.
  3. Reading only the sub-queries that name you. Those confirm you are on the shortlist. The unbranded ones are the ones still in play, and they are where a smaller brand enters an answer.
  4. Comparing engines without normalising. A category looks "more competitive" on Gemini purely because Gemini issues ten sub-queries where Perplexity issues one. Compare shares within an engine, never counts across engines.

In our view fan-out data is the closest thing AI search has to a free intent study, and its value is in the half of each string the engine wrote rather than the half it copied from the prompt. The answer tells you what an engine concluded about your category; the fan-out tells you which questions it felt obliged to settle first, and those are the pages that get read. For the metrics that sit downstream of this, see GEO metrics explained, and for the retrieval step that follows the fan-out, how ChatGPT, Perplexity and Claude choose their sources.

Frequently asked questions

What does fan-out data tell you about buyer intent?

It tells you which conditions an engine decided to verify before it recommended anyone, written in the engine's own words. Seer Interactive's test of 501 prompts on the Gemini 3 API returned 10.7 sub-queries per prompt on average, from 3 to 28, and 26.4% of them contained a brand name, which means a quarter of the retrieval work was already about specific vendors. The sub-queries that name no brand are the category questions you can still win.

Are fan-out queries the same as keywords?

No, and treating them as keywords is the fastest way to misread them. Seer found 95% of Gemini 3 fan-out queries with zero global search volume, and DataForSEO's analysis of 100,000 ChatGPT prompts shows fan-out queries landing in the 61 to 90 character range against 31 to 60 for the prompts that produced them. They are machine-written retrieval strings, so they describe how an engine frames your category rather than how many people type a phrase.

How many times should you collect a prompt before reading its fan-out?

Enough times to see a cluster repeat, which in practice means several runs per week for a month. Surfer's stability test over 1,600 grounded Gemini runs found only about 27% of fan-out queries staying consistent between runs, with 66% appearing exactly once, while about 90% of the fan-outs fell into four or fewer topic clusters. The clusters hold even when the strings do not.

Can you see the fan-out behind Google AI Overviews?

Not on Google's own answer surfaces. The model APIs expose their searches, with one server_tool_use block per search on the Claude API and a google_search_call step on the Gemini API, and DataForSEO returns fan_out_queries for ChatGPT, Gemini, Claude and Perplexity. For AI Overviews and AI Mode, no published field carries the sub-queries, so what you read there is reconstructed rather than reported.

What does a brand name inside a sub-query mean?

It means the engine stopped exploring the category and started checking a specific vendor, so the retrieval that follows points at pages about that vendor. Seer measured 26.4% of Gemini 3 fan-outs carrying a brand name, and Ahrefs found 6.8% of ChatGPT fan-outs using the site: operator, with 75.9% of the most common site: searches pointed at a brand's own site. Both signals say the same thing: at that step, your own pages are the ones being read.