Key takeaways
- A single prompt typically becomes somewhere between 4 and 20 background searches.
- You cannot see the real sub-queries, but you can infer them four different ways.
- Citations in the answer are your best free clue about what was retrieved.
- Cover the sub-queries as sections on one page, not as separate thin pages.
- Prompt-level rank trackers miss this entirely, which is why coverage beats position.
Query fan-out is the step where an AI search system turns your one question into a batch of narrower background searches, retrieves sources for each, then writes a single answer from the results. You cannot read the real sub-query list, but you can reconstruct most of it from the citations in the answer, the follow-up suggestions the interface offers, asking the assistant directly what it searched, and fan-out simulator tools. Then you answer each sub-question as its own section on one page.
A worked example
Take a commercial question a real buyer might type:
"What is the best free AI SEO course for a marketing team?"
The engine does not search that string. It decomposes the intent. A plausible fan-out looks like this:
- best free AI SEO course 2026
- AI SEO course reviews
- free vs paid AI SEO training
- AI SEO curriculum topics covered
- what is AI SEO
- GEO vs AEO vs SEO difference
- AI SEO course for teams vs individuals
- how long does an AI SEO course take
- does an AI SEO certificate matter
- AI SEO course no signup required
Notice what happened. Only the first two resemble the original phrasing. The rest are definitional, comparative, or objection-handling. A page that only targets "best free AI SEO course" competes for two of ten slots. A page that also explains what AI SEO is, how GEO differs from SEO, what the curriculum covers and how long it takes competes for most of them.
That is the whole lesson. Fan-out turns one ranking contest into ten retrieval contests. Breadth of genuine coverage wins more of them than keyword precision does.
How many sub-queries are there?
Published analyses land in a wide band. Google AI Mode is commonly reported around 8 to 12 sub-queries per prompt, while ChatGPT has been observed anywhere from roughly 4 to 20 depending on the question and how much it decides to browse. Some recent observations suggest ChatGPT has tightened toward fewer, broader fan-outs per prompt.
Do not anchor on a number. The useful takeaway is the order of magnitude: one prompt means many searches, not one, and the count moves as these products change. Google does document the fan-out behaviour itself for AI Overviews and AI Mode, so the mechanism is not speculation even though the exact counts are.
Four ways to find your sub-queries
1. Read the citations, the free method
Ask your target question in AI Mode, ChatGPT search, or Perplexity, then study every source it cites. Each citation was retrieved to satisfy some sub-query. Open the cited pages and ask what specific question each one answers. Work backwards and you will recover a good share of the fan-out.
This is the highest signal free method, because it shows what actually got retrieved rather than what a tool guesses.
2. Expand the follow-up suggestions
AI Mode and Perplexity both offer related or follow-up questions after an answer. Those are adjacent to the fan-out tree. Click through two levels and log everything. You are mapping the neighbourhood the system considers part of your topic.
3. Ask the assistant what it searched
In ChatGPT or Perplexity, follow up with something like: "List the exact web searches you ran to answer that." You will often get a usable list. Treat it as a reconstruction rather than a log, since models can rationalise after the fact, but it is fast and frequently accurate enough to act on.
4. Use a fan-out simulator
Several tools now generate likely sub-queries from a seed prompt, including free simulators. They are predictions, not observations, so use them to fill gaps after you have done method one rather than as your only input.
One caution worth repeating: most AEO rank trackers monitor the prompt you typed, not the sub-queries the model ran. That gap is exactly why a page can feel invisible in a tracker while quietly earning citations, and why tracking prompts rather than keywords still misses part of the picture.
Classify before you write
Once you have twenty or thirty candidate sub-queries, sort them into three buckets:
- Must answer. Directly about your topic and your offer. These become dedicated sections with question headings.
- Should answer. Definitional and comparative context. These become short subsections or a comparison table.
- Ignore. Adjacent but off-intent. Covering these dilutes the page.
Then write one section per must-answer item, each opening with a direct two or three sentence answer before any elaboration. That answer-first shape is what makes a passage quotable, which is covered in more depth in how to write citation-worthy content.
Why one page beats ten thin ones
The instinct is to spin up a page per sub-query. Resist it. Fan-out retrieval operates on passages, not whole documents. A single authoritative page with ten well-marked sections offers ten retrievable passages under one URL that accumulates all the authority signals. Ten thin pages split those signals ten ways and each one struggles to be considered at all.
The exception is when a sub-query is genuinely a different topic with its own search demand. That deserves its own page, and then you link the two so the relationship is explicit.
Quick checklist
- Pick one real buyer question, not a keyword.
- Run it in AI Mode, ChatGPT and Perplexity. Log every citation.
- Reverse-engineer the sub-question behind each citation.
- Expand follow-up suggestions two levels deep.
- Ask the assistant what it searched. Note the overlap.
- Classify into must, should and ignore.
- Write one section per must-answer, answer-first, question as the heading.
- Re-run the prompt in two weeks and check whether you are now cited.
Start here: if the mechanism itself is still fuzzy, read query fan-out explained first, then come back to this. The full method is in the free AI SEO course.