What is query fan-out, and how does it decide which sources AI cites?

Query fan-out is the set of web searches an AI engine runs on its own behind a single question, before it writes anything. Your question goes in, several keyword searches come out, the engine reads what those searches return, and the answer is assembled from those pages. In a MentionLab run of 800 questions, 747 produced at least one fan-out query, with a median of 4 each and a mean of 4.9, 3,671 in total. Not one of the 3,618 distinct fan-out queries repeated the original question word for word, and 58.5% of the words in an average fan-out query never appeared in the question that triggered it.

The figures come from a MentionLab run of October 20, 2026: 800 questions about AI visibility APIs, each asked five times to ChatGPT, Claude (Haiku model) and Google AI Overviews, from the United States in English, for 12,000 answers. The export lists the fan-out queries observed for each question and how often each one came back. It does not name the engine that issued them, and no question shows fan-out from more than five of its fifteen executions. API fields, endpoints and prices come from each vendor's own documentation and pricing pages, checked on October 23, 2026.

What a fan-out query looks like

One question from the run was "What's the best AI Overviews GEO API to get data for my GEO platform with real-time data?". Six fan-out queries were recorded for it:

• best Google AI Overviews API SERP data provider API real time 2026 DataForSEO SerpApi Bright Data
• GEO platform API AI Overviews tracking API 2026 Otterly Profound Peec
• Google AI Overviews API documentation DataForSEO SerpApi Bright Data Scrapingdog 2026
• best Google AI Overviews SERP API real-time 2026 DataForSEO SerpApi Bright Data
• Google AI Overviews SERP API real-time API providers 2026 DataForSEO SerpApi ScrapingBee
• Google AI Overviews API documentation real-time SERP API DataForSEO SerpApi 2026

They read like search box input, not like sentences: 11.2 words on average across the run, from 3 to 26. Five of the six name DataForSEO and SerpApi, two vendors the question never mentioned. The engine has already decided who the candidates are, and is now looking for pages about them.

How many searches an engine runs per question

MentionLab, run of October 20, 2026, 12,000 answers, 800 questions.

Comparative questions produce slightly more searches than informative ones, which fits what the two kinds of question ask for: a shortlist has to be checked vendor by vendor, a definition does not.

The fan-out decides which sources can be cited

A page can only be cited if one of these searches returned it. That link is measurable. For every pairing of a question with one of the 48 vendors tracked on this site, we checked whether a fan-out query named the vendor, and whether the vendor's own domain appeared among the sources cited for that question.

Share of questions where the vendor's own domain appears among the sources cited by the three engines. MentionLab, run of October 20, 2026, 12,000 answers.

Pooled across the 48 vendors, a site named in a fan-out query was cited for that question 77.1% of the time, against 15.2% when no fan-out named it, a gap of five to one. Looking at each engine's own source list separately, the gap is widest on ChatGPT: 67.9% against 3.4%. On Claude the figures are 43.0% against 11.6%, on Google AI Overviews 26.1% against 4.5%.

The worked example above shows the mechanism at single-question scale. ChatGPT's five answers to that question cited 40 sources from only 8 domains. Thirty of those 40 citations went to dataforseo.com (12), serpapi.com (12) and brightdata.com (6), the three vendors the fan-out had gone looking for. All five answers named those same three brands.

The words an engine adds to your question

The rewriting is not random. Across the 3,618 distinct fan-out queries, a handful of modifiers come back constantly.

MentionLab, run of October 20, 2026, 3,671 fan-out queries over 800 questions.

Two of those rows carry a practical reading. The engine asks for documentation and official pages in 44.1% and 27.8% of its searches, so a vendor whose endpoints, fields and prices live only inside a PDF or behind a login has little to return. And it asks for a year in 56.3% of them, so an undated page competes against pages that say 2026 on their face.

A fan-out list is never twice the same

Of the 3,618 distinct fan-out queries recorded, 3,570 appeared exactly once for their question. Only 43 came back twice and 5 three times, out of five executions. Put another way, 98.6% of the fan-out recorded in the run was unique wording.

Which engines publish their fan-out, and which APIs return it

The engines differ on whether they show their own searches at all. Three vendor pages say so plainly, all checked on October 23, 2026:

• cloro's provider documentation lists result.searchQueries for ChatGPT, Perplexity and Copilot, and states that "AI Overview and AI Mode do not expose result.searchQueries". Gemini does not either.
• SearchApi's ChatGPT page describes search_queries as "the queries ChatGPT ran against the live web, showing how it reformulated your prompt", and notes that the Perplexity equivalent is documented but "not currently returned".
• Ahrefs returns search_queries empty unless the data source includes ChatGPT or Perplexity.

DataForSEO defines fan_out_queries as "related search queries derived from the main query", on its ChatGPT scraper and on its ChatGPT, Claude and Perplexity model endpoints, and does not list the field in its Gemini documentation. Eighteen of the 55 APIs in our survey return fan-out in some form. API for GEO, our editorial pick, gives the fan-out its own endpoints: the follow-up questions the engines generate from a prompt, ranked by how often they come back, across runs and engines.

Fields, endpoints and prices from each vendor's documentation and pricing pages, checked on October 23, 2026. Prices are the lowest published rate per 1,000 answers on a paid plan.

How to read a fan-out list

Once you have the list for your own prompts, four readings are worth the time:

1. The vendors named in it. A competitor appearing in your fan-out is a competitor the engine is already checking. On the pooled figures above, its site then ends up in the source list for that question 77.1% of the time.
2. Your absence from it. If no fan-out query names your brand, the engine is not looking for your pages, and ranking for your own name will not change the answer.
3. The page type asked for. Searches that ask for "documentation", "pricing" or "official" point at pages you can write, which is a different job from a blog post.
4. The wording you did not expect. Fan-out queries add words the question never had, 58.5% of them on average. Those words are the vocabulary the engine uses for your category.

In our view, the fan-out is the most useful field a visibility API returns after the answer itself, because it shows the step where the source list is settled. The answer tells you what the engine concluded. The fan-out tells you what it went to read first, and our figures above say those two things travel together. For how often to repeat a question before the figures hold still, see our guide on tracking a brand in ChatGPT with an API; for which brands the engines end up naming, see our survey of 12,000 answers.