When someone types a question into an AI answer engine, the system rarely searches for exactly those words. It breaks the question into smaller ones, looks each up, and stitches the findings into one response. If your page answers only the headline question, you may never be considered for the pieces that actually build the answer.
That behavior is called query fan-out, and it changes how AI SEO content should be planned.
What Query Fan-Out Is
Google describes it directly. In its AI Mode announcement, Google says the technique splits a complex question into related subtopics and runs several searches at once, drawing on web content, the Knowledge Graph and shopping data, then combines the results into one response. Google’s documentation for site owners adds that AI Overviews and AI Mode may run several related searches covering different subtopics and sources, so the supporting links can be broader and more varied than in a classic results page.
Two limits matter. Google does not publish which sub-queries it generates for a given search, so anyone claiming to know them is inferring. And Google says there are no extra technical requirements: a page must be indexed and eligible to show a snippet, and even then nothing is guaranteed.
An Illustration
Take a search such as “best accounting software for a ten-person design studio.” The following sub-questions are my own illustration of how such a request could decompose, not Google’s actual output:
- What does it cost per user at ten seats?
- Does it handle project-based billing?
- How hard is moving data from spreadsheets?
- What do small studios say about support?
- Which alternatives are cheaper?
A page titled “Best Accounting Software” that lists ten products with generic blurbs answers none of these well. A page that handles pricing at a specific team size, migration effort and billing models gives a system several distinct things to pick up.
What the Early Studies Suggest
Evidence is thin, mostly observational and published by vendors, so treat it as directional.
AirOps analyzed 16,851 queries in ChatGPT, run three times each. Most queries (88.6%) produced exactly two sub-queries, while 8.8% produced none, typically simple product or entity lookups. Pages whose headings closely matched the query were cited more often, with citation rates rising from roughly 30% to 41% across similarity buckets. Covering more of the fan-out did not keep helping: moderate coverage outperformed full coverage once query match was held constant. Retrieval rank remained the strongest predictor. The study covered only ChatGPT, and AirOps sells related software.
Grow and Convert tested 100 buying-intent prompts, 83 of which triggered a search. Only about 40% of cited sources appeared in the top ten results of Google or Bing for the visible fan-out queries. The authors’ explanations for the rest, such as hidden queries or training data, are theories. Their sample is small, and they are an agency selling in this space.
Together these suggest two things: ranking for the sub-question still matters, and a page can be cited without ranking where you would expect. Neither study proves that any tactic causes citation.
The Sub-Question Map
Instead of starting from keywords, start from the decisions behind the main question. This takes about an hour per page.
- Write the main question in the customer’s words, not yours.
- List the decisions inside it. Most fall into six types, shown below.
- Gather real phrasing from sales calls, support tickets, reviews and your Search Console queries.
- Assign each sub-question a home, either a section on this page or its own page.
- Answer it first, then support it. Put a direct two-to-four sentence answer under a heading that mirrors the question.
One Page or Several?
A sub-question deserves its own page when it has its own audience and could stand alone, such as “how to migrate from spreadsheets.” Keep it as a section when it only makes sense inside the main decision, such as per-seat pricing for one product. Splitting everything creates thin, competing pages; merging everything creates a page nobody can scan. Link related pages with descriptive anchor text.
Mistakes to Avoid
Chasing completeness. The AirOps finding on moderate coverage is a reminder that a page answering every conceivable question is not automatically better. Answer the ones that matter to the decision.
Question-stuffed headings. Five phrasings of one question read as padding. Pick the clearest.
Treating a tool’s guesses as fact. Fan-out simulators approximate the process with model prompts. Use them for brainstorming, not measurement.
Stale answers. A year-old pricing answer can be worse than none. Date it and revise it.
How to Test a Page
Write the ten prompts a buyer might type. Run each in two or three assistants and in Google’s AI features, then ask a simple question: which parts of the answer came from competitors, and which sub-questions did my page leave unanswered? Expect answers to vary between runs, so look for patterns, not single results. Sarkar SEO runs this kind of gap review before drafting, because it shows which questions to answer rather than which keywords to repeat.
Putting It Into Practice
Choose one commercial page that already ranks but underperforms. Build its sub-question map, add the missing answers and update the figures. Sarkar SEO suggests changing one page at a time so you can attribute results, then repeating the process across the pages that drive revenue.
The shift is simple to state and hard to do well: stop writing for a phrase and start writing for a decision. If you would like help mapping the questions behind your most valuable searches, Sarkar SEO can build the map and the content plan that follows from it.
