Retrieval reference · Published September 2, 2026
Query fan-out
Query fan-out is the set of underlying searches an answer engine generates from one user prompt before it composes an answer. The prompt is the request. The fan-outs reveal what the engine decided it needed to retrieve.
A prompt is not a keyword with extra words
An answer engine can decide whether to search, rewrite the request, add a year or location, branch into comparison questions, and consult several result sets. Those branches create the candidate pages available to the answer.
This changes the unit of analysis. Matching the exact prompt wording is not enough if the retrieval system searches for something else. A page can answer the original question and still never enter the candidate set.
Fan-out is diagnostic, not a content-production queue. The useful question is which branch created the gap and which source is best suited to close it—not how quickly a team can create a page for every generated string.
ChatGPT usually issued two or three searches
Measured sample36.4%
Two searches
52.9%
Three searches
10.7%
One, four, or five
In the published ChatGPT fan-out sample, 36.4% of prompts produced two searches and 52.9% produced three. Together, 89.3% produced two or three underlying searches. The remaining 10.7% produced one, four, or five.
The public study does not disclose the fan-out sample size. These percentages describe the measured prompt mix and should not be presented as a universal ChatGPT rate.
Read the 250M-response study →Fan-out and Google overlapped, but they were not the same
A paired analysis of 1,000 Google result pages and 1,000 ChatGPT executions found 39% overlap between ChatGPT's fan-out results and Google's results. The study supports a connection between search visibility and the candidate set. It does not show that copying a fan-out into a title causes a citation.
The practical use is to inspect the branches. If an engine searched a comparison the site does not answer, that is a content gap. If the right page ranked but the engine selected another passage, the problem sits later in the path. Retrieval and citation selection are related, not interchangeable.
Claude did not fan out like ChatGPT
Keep engines separate| Measure | Claude | ChatGPT | Reading rule |
|---|---|---|---|
| Fan-outs containing a year | 94% | 17% | Measured with 2025 and 2026 in the observed query strings; sample size is unpublished. |
| Repeated fan-out strings | ~65% | Not reported | The repetition count and matching rule are not public. |
| Citation-domain overlap | 8% with ChatGPT | 8% with Claude | Average domain overlap across more than 600 queries; the public article does not publish the formula. |
In a separate alignment analysis, adding a year to the page title increased query similarity by 17%. That is an observed relationship in the measured set, not an instruction to add the current year to every title.
Read the State of AEO study →Fan-out changes over time
In a separate 2026 follow-up, the share of measured ChatGPT fan-outs explicitly adding Reddit rose from 0.15% in January to 3.68% by late May. The analysis used roughly seven million recent citations and a different time window from the original Reddit study.
That movement is why fan-out language belongs in the measurement record. A page can stay unchanged while the engine changes how it researches the same prompt. The follow-up should not be pooled with the 2025 six-engine Reddit aggregate.
Read the Reddit citation study →How to inspect a fan-out gap
Six steps- 01
Confirm that the engine searched
A live page cannot enter the retrieval path when the answer was generated without web search. Record the product mode and search trigger first.
- 02
Save the underlying query strings
Keep each fan-out exactly as issued, including dates, comparison language, locations, and other terms the engine added.
- 03
Inspect the candidate results
Run the fan-out strings through the relevant search index when that index is known. Compare domains, URLs, page types, and passages—not only rank.
- 04
Trace the selected evidence
Record which candidate pages were cited, which claim each citation supports, and whether the brand was included accurately in the final answer.
- 05
Change the smallest useful source
Improve an existing page, publish a focused source, correct product information, or earn corroboration where the engine already looks. Do not create one page per fan-out.
- 06
Rerun the same prompt
Keep engine, market, account state, and other filters fixed. A different fan-out can explain movement even when the page did not change.
Limits of the public record
The published fan-out findings come from separate analyses with different units and subsamples. The ChatGPT fan-out count, Google overlap, Claude year insertion, and Reddit trend are not one pooled experiment.
The public material does not include the ChatGPT fan-out denominator, Claude fan-out sample size, overlap formula, or raw query logs. Those omissions limit the claims this page can make. It reports the observations and the workflow they inform without inventing missing precision.