Traditional keyword research starts with a tool, a seed term, and a volume column. AI search breaks that routine in two ways. People type long, conversational questions that rarely appear in keyword databases, and AI SEO systems often rewrite those questions into several hidden searches before answering. If your research only covers the visible keyword, you are missing most of what shapes the answer. This article shows how to find the questions that matter and turn them into a repeatable research process.
Why keyword lists fall short
The unit of research has changed. One comparison describes classic keyword research as covering two- to five-word search terms, while prompt research deals with conversational questions of 15 to 25 words, with far less volume data available for the latter. User behavior supports the shift. A 2025 Microsoft Research study found that people using LLM-based search finished tasks faster with fewer, more complex queries than traditional search users. similarwebtryvizup
So the buyer’s real question may be “which project management tool works for a ten-person agency with client approvals and a small budget,” not “project management software.” A page built only for the short term won’t obviously answer the long one.
What query fan-out actually is
When a chatbot with web search turns on, it may not search your exact words. OpenAI’s documentation says ChatGPT typically rewrites a query into one or more targeted queries and sends them to search engines. Those rewrites are known as fan-out queries, and they are not displayed to the user.
Two data points show the scale. DataForSEO analyzed 100,000 ChatGPT prompts and found that nearly half triggered fan-out queries. Seer Interactive studied Gemini 3 and reported an average of 10.7 fan-out queries per prompt, with 95% having zero monthly search volume and an average length of 6.7 words. The same study found 21.3% of fan-outs contained a year and 26.4% included a brand name. Fan-Out Queries Featured Image +2
Treat these as snapshots from vendors with a commercial interest in AI visibility tools. Seer itself raised an open question: whether fan-outs are consistent enough to act on or different every time. That honesty is useful. Fan-outs are a research lens, not a rulebook.
Where to find prompts you can trust
AI platforms do not publish query data the way Google Search Console does, so you must triangulate.
Search Console long queries. Your own data is the closest public proxy. One practitioner suggests filtering to queries of ten words or more using the regex ^(\S+\s){9,}\S+$ under Performance, then Query, then Custom (regex), while noting the right threshold depends on your category. Start there, then try shorter thresholds.
People Also Ask and related questions. These show real queries entered into Google, and they often read like conversational prompts.
Your own conversations. This is the source competitors can’t copy. Sales call notes, support tickets, chat transcripts, and onboarding questions contain the exact wording and constraints your buyers use. We consider this the richest source at Sarkar SEO, and it costs nothing.
Community discussions. Forums and review sites show how people phrase doubts, comparisons, and objections.
AI-expanded variations. You can ask an LLM to generate prompt variations, but label them as synthetic. They show plausible phrasing, not proven demand, so validate them against real sources.
Fan-out tools. Some AI visibility platforms now expose the sub-searches models run. One suggests looking for fan-out queries that recur across many prompts as the searches models rely on most in your category. Recurrence is a sensible filter, since one-off variants may be noise.
The Anchor, Fan, Fill method
This is a practical process we use at Sarkar SEO. It is our own approach rather than a standard.
- Choose an anchor. Pick a topic tied to revenue, such as “payroll software for small restaurants.”
- Write the fan. List the sub-questions an AI would need to resolve before answering. Cover seven angles: definition, comparison, constraint (budget, size, location), price, risk or compliance, how-to, and alternatives. Add the current year or a brand name where it naturally fits, since the Seer data shows both appear often.
- Gather real wording. Match each angle against Search Console, Questions in your own inbox, and community threads.
- Test it. Run the anchor prompt in two or three AI tools with search on. Record which sources are cited and which angles they cover.
- Map your coverage. Build a simple grid: rows are fan-out angles, columns are your pages. Mark each cell as covered well, covered thinly, or missing.
- Fill the gaps. Add a clear, specific answer to the existing page where it fits naturally. Only create a new page when the question deserves its own depth.
Prioritize ruthlessly
Ten sub-queries per prompt does not mean ten new pages. Seer asked whether a keyword list suddenly grew tenfold, and the sensible answer is no. Score each gap on three points: does it relate to a purchase decision, can you add something others lack (data, experience, a tool), and does it recur when you rerun the test three times? Work on gaps that satisfy all three first.
Common mistakes
- Treating zero-volume as zero value. Seer’s finding that 95% of fan-outs had no search volume means the usual filter would discard nearly all of them.
- Trusting synthetic prompts without checking real sources.
- Building a thin page for every fan-out variant.
- Running a single test and calling it data.
- Forgetting the buyer. A question list is only useful if it maps to what someone needs to decide.
