The short answer
When someone asks ChatGPT, Gemini, or Claude a question that needs current information, the assistant does not search for the question as typed. It rewrites it into several shorter searches, often adding words like "best", "reviews", "vs", and the current year, then combines what it finds. Pages that answer several of those searches together get cited most.
This matters because the question your customer typed and the searches that decide which businesses appear in the answer are different things. Most content is still written for the first. This piece is about the second.
What is query fan-out?
Query fan-out is the technique AI search systems use to break one question into multiple related sub-queries, run them at once, and merge the results into a single answer. Google named the technique when it introduced AI Mode in 2025, and ChatGPT's search tool, Perplexity, and other assistants follow the same basic pattern.
A prompt to an assistant is long and specific. Semrush's analysis of ChatGPT prompts found that only around a third resemble traditional search queries; the rest are situational requests with context no keyword tool would record. The assistant cannot run that paragraph as a search, so it extracts the searchable parts, adds what it thinks will produce useful results, and fans out.
A customer's prompt such as "I run a small clothing brand in Jaipur, should I use Shopify or get a custom website built, and what will it cost?" might become searches like these:
| Likely sub-query | What it is really looking for |
|---|---|
| shopify vs custom website clothing brand | A direct comparison |
| shopify pricing india 2026 | Current, dated costs |
| best ecommerce platform for clothing brand india | Evaluations and lists |
| custom ecommerce website cost india | Price ranges from builders |
| shopify india transaction fees reviews | Downsides and experiences |
None of those is the question the customer asked. All of them decide which pages the assistant reads before answering.
How many searches ChatGPT runs per question
The number of searches depends on the model and the question. A study by Peec AI of 5 million fan-out queries collected in April 2026 measured an average of 2.1 searches per ChatGPT response, 1.4 for Perplexity, and 6.8 for Grok. Later analyses after newer ChatGPT default models rolled out found substantially more.
One analysis published in August 2026 found the share of ChatGPT prompts producing only a single search fell from 94 percent to 43.5 percent after a new default model, and reasoning modes routinely run many rounds of searching for one answer. The direction is clear even though the numbers move with every model update: assistants are searching more per question, which means more chances for a page to be found and more competition for the final citations.
These figures come from third-party measurement and will date quickly. The structural point, that one question becomes several searches, is what to plan around.
The words AI adds to your customer's question
AI assistants consistently inject evaluative and freshness words that the user never typed. In the April 2026 study, the most frequently added words in ChatGPT's searches were "best" (15.33 percent), "what" (8.72 percent), "reviews" (6.84 percent), "2026" (5.44 percent), "top", "comparison", "vs", "company", "services", and "software".
For advice-seeking questions, "best" was added about one time in four. Read together, the list describes what the assistant believes a good answer needs: a ranked or evaluated set of options, independent opinion, a comparison, and current information. That is a useful brief for any commercial page.
- "Best" and "top" mean the assistant is looking for evaluated options, not a single vendor's claims.
- "Reviews" sends it towards third-party opinion, so what others say about you matters as much as what you say.
- "Vs" and "comparison" reward pages that compare options honestly, including ones you do not sell.
- The current year favours content that is genuinely current and says when it was updated.
Where AI assistants look: trusted sites, reviews, and Reddit
AI assistants increasingly target sources they already trust. Analyses of ChatGPT's searches in 2026 found growing use of site-restricted searches aimed at specific domains, with official and government sites favoured for legal and medical facts, manufacturers for specifications, and review platforms and Reddit for opinions and experiences.
References to Reddit in ChatGPT's fan-out queries were measured rising from around 0.15 percent to 3.68 percent between January and May 2026. Meanwhile, the number of distinct sites ChatGPT cited per answer fell after its March 2026 update, from about 19 to 15 in one analysis, even as it read more pages. More reading and fewer citations means the pages that do get cited are the ones that answer most completely.
For a business, the implication is that your own site is necessary but not sufficient. Consistent details on review platforms, directories, and the communities your buyers use shape what the assistant finds when it looks beyond you. Our guide to tracking brand visibility in AI search covers how to monitor this.
Why pages that cover several angles get cited
ChatGPT combines results from its sub-queries using a method called reciprocal rank fusion, which scores a page higher when it appears in the results of several different searches. A page that ranks for the comparison, the cost, and the reviews query outscores pages that rank for only one, even if they rank higher there.
That is the single most practical insight in the fan-out research. It rewards depth within a topic rather than volume across keywords. A page that explains what something is, what it costs now, how the options compare, and what the drawbacks are, each in a clearly headed section, has more ways to be retrieved than five short pages each covering one of those.
How to map the fan-out for your own business
To map fan-out for your business, write the full questions buyers would ask an assistant, predict the sub-queries using the injected words above, then check which of those your pages answer clearly. Test real prompts in ChatGPT, Gemini, Claude, and Perplexity and record which sources are cited. The gaps between the two lists are your content plan.
- Collect real questions. Sales calls, enquiry forms, WhatsApp chats, and reviews contain the exact context buyers bring to assistants.
- Predict the sub-queries. For each question, write the category plus "best" or "top", plus "reviews", plus the current year and "cost" or "pricing", and the likely "X vs Y" comparisons.
- Audit your pages. For every sub-query, find the section on your site that answers it in the first two sentences under a heading. If there is none, that is a gap.
- Test in the assistants. Run the full prompts, note who gets cited, and read those pages. They show what the assistant considered a complete answer.
- Repeat quarterly. Models change how they search several times a year.
How to structure a page so it matches sub-queries
Structure pages around the questions inside the question. Give each sub-topic, such as cost, comparison, drawbacks, and how to choose, its own descriptive heading, then answer it in a self-contained paragraph of 40 to 60 words directly beneath. Add tables for comparisons and prices, FAQs for follow-ups, and an honest updated date.
- Headings that mirror sub-queries. "What does X cost in 2026" retrieves better than "Let's talk numbers".
- Answer first. The opening sentences under each heading should stand alone if quoted, because that is how assistants extract them.
- Tables for comparisons and prices. Structured comparisons are easy to extract and are exactly what "vs" and "comparison" searches look for.
- Structured data. Article, FAQ, product, and organisation markup help systems understand the page; see our guide to schema markup for AI search.
- Evidence and dates. Named sources, specific figures, and a genuine revision date matter more when the assistant is adding the year to its search.
This blog follows that structure: each section opens with an answer that makes sense on its own. The broader practice is covered in our guide to generative engine optimisation.
What not to do
Do not create a thin page for every sub-query, stuff "best" and the current year into titles, or change dates without changing content. Separate pages for each variation compete with each other, keyword-stuffed titles read as low quality to both people and ranking systems, and false freshness is detectable and erodes trust.
Equally, do not ignore ordinary search while chasing assistants. Fan-out queries are still run against search indexes, so pages that are not crawlable, fast, and indexed are not in the pool the assistant draws from. And keep volatile facts current, because a page quoting last year's prices is exactly what a search containing "2026" filters out; our piece on content decay and refreshing old pages explains how to prioritise. If you would like us to map the fan-out for your own buyers' questions, get in touch.