How does query fan-out decide which pages an AI engine finds?
Query fan-out decides which pages an engine finds: it rewrites one question into sub-queries, and in an October 2026 preprint every page ChatGPT cited had come back from its own search on those rewrites. Check the sub-queries your buyers' questions produce, then give each one a specific page that matches.
Does a page have to be retrieved before an engine cites it?
Chen and Koudas, a University of Toronto preprint covering 158,099 answers from API models, found that every page ChatGPT cited had come back from the engine's own search on sub-queries the model wrote for that answer. Where a page ranked in that search explained 32% of which pages from one site got cited (Chen and Koudas). All page features combined explained 2.8% (Chen and Koudas).
The authors advise a GEO vendor, and the work is a preprint.
Jiang et al.1, a preprint covering four ChatGPT Web settings, saw the same direction: candidates retrieved more often across repeated runs were cited more often. Their URL-level check found Google Search and OpenAI's Responses API each recovered only about one quarter of the URLs ChatGPT Web cited, even combined.
Our read: a page the engine's own rewritten searches never return has no route to a citation, so retrieval is the first gate. The chart sets the figures side by side.
How many sub-queries does an engine run per question?
Fan-out is not a fixed multiplier. It varies by engine, by model version and by whether the engine searches at all. The table lists four measurements, each with its source and date.
| Source | Date | What it measured |
|---|---|---|
| Ahrefs2 | Aug to Sep 2026 | ChatGPT searched the web for 14.4% of 66.6 million responses, with 1.6 fan-out queries per search |
| Profound3 (Sonnet 5.5 only, version-specific) | Oct 2026 | 1.17 fan-outs per searched answer against 1.83 for Sonnet 5, across 28,440 runs |
| Profound (Sonnet 5.5 only, version-specific) | Oct 2026 | 7.7% of 711 prompts switched between searching and not searching, against 29.1% for Sonnet 5 |
| Jiang et al., preprint | Oct 2026 | GPT-5.4 Thinking often ran several retrieval rounds, while Instant settings stayed mostly single-round |
Profound's own data also links the lower fan-out of Sonnet 5.5 to a citation candidate pool that is 36% smaller than Sonnet 5's. The Sonnet rows describe those versions only, not Claude in general.
What do the rewritten queries ask for?
Ahrefs found in August and September 2026 that 75.9% of ChatGPT's site: searches targeted brand sites. English fan-out queries containing 'official' moved from 10.1% in August to 13.1% one month later.
For factual questions about pricing, product features or company details, Ahrefs found at least one official source cited in 84.4% of cases, across 20,000 US prompts.
Profound's own data shows Sonnet 5.5 (a version-specific finding) wrote '2026' into 59% of its fan-outs, against 26% for Sonnet 5. Its most-cited pages carried the year in the title at 49.96%, against 37.97% for Sonnet 5, across 28,440 runs in October 2026. The median age of its top 10% of cited pages was 107 days, against 164 for Sonnet 5. The chart shows the year figure.
Chen and Koudas found that citation was predicted by title match and by paragraph match with the question. Within the same site, pages with FAQ blocks, structured data, summary boxes or fast loading were not cited more often in the study.
How do you see which fan-out queries miss your pages?
Bing's Topics Portfolio, announced on 13 October 2026, groups grounding queries into thematic clusters. For each topic it shows an approximate size, the number of sites cited and the number of distinct queries.
Bing Webmaster Tools also lists Top Opportunity Queries: queries in topics where other sites appear but yours does not. They come as citation ranges, and Bing labels every number a directional signal, not a precise ledger.
Bing says topic labels are AI-generated and imperfect. Topics with too little site presence or ecosystem activity are hidden, so smaller sites may not see the feature. Branded queries, including a competitor's, can also show up as opportunities. These are platform statements about Bing's own tool.
For ChatGPT, Chen and Koudas found its own search on rewritten sub-queries rarely coincides with Google's or Bing's results. Jiang et al. found external proxies matched ChatGPT Web better at hostname and entity level than at exact URL. The diagram shows the path that any fan-out tool is trying to observe.
Which pages does each engine find for the same question?
New Map's measurement4 found Gemini citing 9.3 sources per answer, 629 sources across 68 answers. In that reading Perplexity gave 0.3 per answer, 21 sources across 68 answers. That is one week's reading of 68 questions about AI visibility tools on 3 engines (146 answers), not a trend: Perplexity cited a source in only 7 of its 68 answers, and Claude Code's 10 answers are too few to read alone. How New Map measures explains the method.
The domain cited in the most answers was reddit.com, in 20 of 146 answers (13.7%) (New Map measurement). Wikipedia.org appeared in 1 of 146 (0.7%). New Map's measurement did not compare which URLs each engine cited for the same question, so overlap between our three engines is not measured here.
Omnia data via Kevin Indig found in May 2026 that 2.37% of cited URLs appeared in all of ChatGPT, Perplexity and AI Overviews for the same prompt, across 3.7M citations.
The chart shows how differently three engines cite for the same questions.
What to do
- Our read from Ahrefs' October 2026 fan-out data: buyers' questions produce sub-queries such as official, price and review, so check whether one page on your site answers each of them plainly.
- Put the answer in the page title and the first paragraph, because title and paragraph match predicted citation in Chen and Koudas. Follow the steps in how to get your brand recommended by ChatGPT.
- If your site is large enough to see Bing's Topics Portfolio, read the Top Opportunity Queries. Then compare them with each engine's real answers, and see plans and pricing if you want that done for you.
Frequently asked
What is query fan-out?
Query fan-out is an engine rewriting one question into several sub-queries and searching each. Ahrefs measured an average of 1.6 fan-out queries per ChatGPT search across 66.6 million responses in August and September 2026.
Does page formatting decide whether a page is cited?
Title match and paragraph match predicted citation in the Chen and Koudas study, while structured data, summary boxes, page speed and FAQ blocks showed no within-site effect. Retrieval rank explained 32% of variation, all page features 2.8% (preprint, October 2026) (Chen and Koudas).
Do all AI engines fan out the same way?
No. Jiang et al. found GPT-5.4 Thinking often ran several retrieval rounds while Instant settings stayed mostly single-round. Profound found Sonnet 5.5 ran 1.17 fan-outs per searched answer against 1.83 for Sonnet 5 (October 2026).
Sources
- Jiang et al. (arXiv), Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations, arxiv.org, 11 October 2026
- Ahrefs, How ChatGPT Does Query Fan-Out (15 Million Queries Studied), ahrefs.com, 12 October 2026
- Profound, What’s changed with Sonnet 5.5, tryprofound.com, 13 October 2026
- New Map measurement, How we measure, 19 October 2026
- Microsoft Bing, Introducing Topics Portfolio (Beta) in AI Performance, blogs.bing.com, 13 October 2026
- Chen et al. (arXiv), What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO), arxiv.org, 10 October 2026
See where you stand
Book a call and find out how visible your brand is across the AI engines your customers already ask.
Book a callFor agencies: See the agency package