Query fan-out is when an AI search system splits one question into several web searches, reads the results, and writes one answer from them. Google documents it for AI Overviews and AI Mode: both "may use a 'query fan-out' technique", issuing related searches across subtopics and data sources (Google Search Central, checked 14 Oct 2026). The page a buyer finally sees cited may rank for one of those searches, not for the question they typed.
OpenAI draws a second line that matters here. Its developer guide for web search says the sources field lists every URL the model consulted, and that this list is often longer than the citations shown in the answer (OpenAI, Web search, checked 14 Oct 2026). So a search can bring your page into the reading list without your page being cited.
This guide shows the real searches ChatGPT ran for 160 shopping questions asked from India, what they have in common, and how to write pages that answer them.
What is query fan-out in AI search?
Query fan-out is the step between a question and an answer. The AI rewrites the question into one or more search queries, sends them to a search index, and reads what comes back. Google says AI Mode does this by "breaking down your question into subtopics" and searching for them at once (Google blog, 20 May 2025, checked 14 Oct 2026).
The same Google post says Deep Search, a longer research mode in AI Mode, "can issue hundreds of searches" for one question. Google also says AI Mode and AI Overviews may use different models and techniques, so their links can differ (Google Search Central, checked 14 Oct 2026).
For a brand, the point is simple. The engine searches for its own version of the question, and the pages that answer that version are the ones it reads.
Which searches did ChatGPT run for real shopping questions?
ChatGPT reports the searches it ran, and they are short, specific and full of buying details. The examples below come from the IndicGEO October 2026 re-run, the second wave of our Hinglish vs English AI shopping study. It asked 10 buying questions in skincare and fashion, in English and in Hinglish (Hindi and English typed in English letters), to ChatGPT with web search, from India, 8 times each, on 13 October 2026. That gives 160 ChatGPT answers, each with its searches recorded. The searches are quoted exactly as ChatGPT ran them; Hinglish questions are described in English.
| Buying question | Asked in | Search ChatGPT ran |
|---|---|---|
| Which brand's jeans should I buy for men in India under ₹2000? | English | best men's jeans brands India under ₹2000 2026 price quality Levi's Wrangler Lee Spykar Pepe Jeans |
| Which sarees should I buy online in India for a wedding on a budget? | English | Two searches: "best affordable wedding sarees online India under 3000 2026" and "trusted online saree stores India return policy budget wedding sarees" |
| Which moisturizer should I buy for dry skin in Indian winters? | English | best moisturizer for dry skin winter India dermatologist ceramides glycerin petrolatum recommendations |
| The same budget wedding saree question | Hinglish | best online saree shopping India budget wedding sarees Myntra Meesho Amazon Flipkart 2026 |
| A face wash question: oily skin, under ₹500 | Hinglish | India best face wash for oily skin under 500 rupees 2026 prices salicylic acid cleanser |
What do these searches have in common?
Across all 160 answers, the searches follow a clear pattern. ChatGPT keeps the market and price the shopper gave, and adds the year and the kind of proof a careful shopper would look for. Our counts, from the same dataset:
- Usually one search, sometimes two: 158 of the 160 answers ran one search. Two ran two searches, both for English questions.
- The market and the year: 160 of 160 searches contain "India", and 151 of 160 contain "2026". ChatGPT looks for current, local pages.
- Expert proof for skincare: 58 of the 80 skincare searches contain "dermatologist". Many also name ingredients, such as salicylic acid, niacinamide or ceramides.
- Brand names for fashion: 34 of the 80 fashion searches name at least one brand. None of the fashion questions named a brand, so ChatGPT added the brand names itself.
- Marketplaces for sarees: 12 of the 16 saree searches name a marketplace, such as Myntra, Ajio, Amazon or Meesho.
- Hinglish questions, English searches: all 80 searches behind Hinglish questions were written in English. So a Hinglish question can lead the AI to pages written in English.
Why does query fan-out matter for a brand?
Fan-out decides which pages the AI reads, and the AI can only cite or name what it read. If your page does not answer the search ChatGPT runs, it is less likely to be in the reading list at all.
Three things follow from the data. First, a page that states its price, its market and the current year matches the searches better than a page that does not. Second, if a search already names three rival brands, the results it brings back are pages about those brands, so being on the comparison pages and lists that rank for that search matters. Third, a marketplace search sends the AI to marketplace pages, so a missing listing is a gap the AI can see. Our marketplace intelligence page explains how to find those gaps.
A page can also be read but not cited, as the OpenAI guide above shows. A page that buries its answer may be read and passed over. For that step, read how AI answers are assembled.
How should a brand use fan-out queries?
Treat each search as a sub-question your site should answer, in a sentence an engine can lift. A practical order:
- 1. Collect the searches. Note the searches behind the buying questions in your category, not only the questions themselves.
- 2. Group them by sub-question. Price band, ingredient or material, use case, market, proof, and where to buy are the usual groups.
- 3. Answer each group on a page. Put the direct answer in the first paragraph, with the price in rupees, the market and the date. Give each sub-question a clear, direct answer, on its own page where it needs one.
- 4. State the proof the search asks for. If the searches say "dermatologist" or "reviews", show the expert input or the review source on the page, with names and dates.
- 5. Check where the searches lead. If they name marketplaces or rival brands, make sure your listings exist and that you are on the lists those searches return.
- 6. Keep the page eligible. Google says a supporting link must be indexed and eligible to show with a snippet, and needs no special files, markup or schema (Google Search Central, checked 14 Oct 2026).
How can I see the searches ChatGPT runs for my category?
Depra AI stores the web searches ChatGPT ran behind each answer it tracks. A single ChatGPT answer in the app shows the products ChatGPT showed and the web searches it ran. On the Depra AI MCP server, the get_chat tool also returns that answer's citations. On the Depra AI MCP server, the list_search_queries tool lists every search with the number of answers that ran it, filtered by date, language, market, topic or tag.
Only ChatGPT reports its searches. Gemini, Perplexity and Google AI Overviews answers are tracked for brand mentions and cited sources, not for searches. Depra AI also does not track rankings inside a marketplace's own search, so a marketplace named in a fan-out search tells you where the AI looks, not where you rank on that marketplace.
To see which engines name you today, run the free AI audit. To track the searches over time, see Depra AI pricing, or book a demo, a free 15-minute call where the Depra AI team sets up your project with you.
What does query fan-out not change?
Fan-out changes which searches run. It does not change the basics that decide whether a page can be found.
- No special file or schema: Google says no AI text files, new markup or special schema.org data are needed for AI Overviews or AI Mode (Google Search Central, checked 14 Oct 2026). An llms.txt file is a separate topic.
- Crawler access still counts: an engine cannot read a page its crawler is blocked from. Check access with the AI crawler checker.
- No guarantee: answering the searches can make a page easier to find and cite. No one can promise a mention, and Depra AI does not.