Prompt Research: Finding the Questions Buyers Ask AI

No tool shows real prompt volumes. OpenAI, Google, Anthropic and Perplexity do not publish how often people ask a given question. Any prompt tool that shows a volume number is estimating, usually by borrowing search keyword data.

The wording problem goes deeper. OpenAI says ChatGPT search typically rewrites a user's question into one or more targeted queries before sending them to search providers. Google describes a query fan-out technique for AI Mode, issuing several related searches across subtopics. The Gemini API generates its own search queries when grounding is on. So the prompt a buyer types is not the query that fetches pages.

What that means: you are not trying to guess an exact string. You are trying to cover the real jobs, worries and comparisons your buyers bring to an AI app. Topical fit matters more than phrasing. The 252,000-trial factorial study found topical relevance and list position were the biggest drivers of being cited first, and Google's AI optimization guide says you don't need to chase every long-tail variation.

The next move is to source prompts from places where buyers already speak in their own words. If you want the difference spelled out, see keywords vs prompts in AI search.

Each source shows a different slice of buyer intent. Use all four if you can.

Sales calls

Listen for the question a prospect asked before they booked the call. "We looked at three tools and couldn't tell which one handles multi-location billing" is a prompt waiting to be written.

Support tickets

Tickets show how customers describe problems before they know your product's vocabulary. Those descriptions are close to what someone types into an AI app.

Search queries

Google Search Console's performance report shows the queries that earned impressions. Its generative AI report, as of October 2026, shows impressions by page, country, date and device, not by query. Use the classic query list as a phrasing source, not as AI demand.

Grounding queries

Bing Webmaster Tools' AI Performance report shows grounding queries: the key phrases the AI used when retrieving content it cited in Copilot, Bing AI summaries and some partner integrations. Microsoft labels this a sample. When the evidence behind this page was checked in August 2026, the report had no API, so you export it by hand.

Plan on an afternoon for the first pass.

1. List the buying jobs. Write the five to ten jobs a buyer hires a product like yours to do. Example, clearly hypothetical: say you sell scheduling software to dental clinics. Jobs might include "reduce no-shows", "book across three locations" and "replace a paper diary".
2. Pull raw phrases. Take the last few months of sales notes, support tickets and Search Console queries. Copy the buyer's sentences verbatim. Do not clean them up yet.
3. Export grounding queries. If your site is verified in Bing Webmaster Tools, export the grounding queries. Tag each one with the job it matches.
4. Write prompts as a buyer would ask an AI. Turn phrases into full questions. Keep the buyer's words and constraints: budget, location, team size, integration. "What scheduling software works for a dental group with three clinics and reduces no-shows?" beats "best dental scheduling software".
5. Cover the buying stages. For each job, write at least one problem question, one shortlist question ("which tools do X") and one comparison question ("X vs Y for Z").
6. Remove your brand name from most prompts. A prompt that names you tests recall of your facts, which is useful but different. Keep a small, separate set of branded prompts for accuracy checks.
7. Test under fixed conditions. Run prompts logged out, or with memory off. OpenAI says ChatGPT may use saved memories when rewriting a search query, which makes logged-in results hard to repeat.
8. Record who gets named and what is said. Note which brands appear, which sources are cited, and any wrong facts about you.
9. Freeze a core set. Pick the prompts that best represent your buyers and keep them unchanged between runs so trends mean something.

If you want a no-cost way to run steps one to five, the free prompt research walkthrough covers it. For whether to write prompts yourself or let software draft them, see own prompts vs tool-generated.

Run your draft list against this before you test anything.

• Every prompt traces back to a real source: a call, a ticket, a search query or a grounding query.
• Every buying job has problem, shortlist and comparison prompts.
• Most prompts do not contain your brand name.
• Prompts include the constraints real buyers mention (size, budget, region, integrations).
• Competitor names appear in comparison prompts the way buyers actually phrase them.
• No prompt is a keyword fragment pasted from a keyword tool.
• No prompt is written to flatter your product's strongest feature.
• Branded accuracy prompts cover the facts AI most often gets wrong: pricing, plans, location, what the product does.
• A frozen core set is marked, with the date it was frozen.
• Test conditions are written down: logged out or memory off, which engines, which date.

If a prompt fails two or more items, rewrite it or drop it. Keyword tools are fine for ideas, but treat them as ideas, not demand data.

A good prompt set gives you answers you did not expect. That is the point of testing.

Signs the research is working

• Answers name competitors you know, and some you did not.
• Different engines give different shortlists for the same prompt. That is normal, because each engine draws on its own index and retrieval.
• You can see which third-party pages engines cite for your category. Being named on those pages is associated with being named in answers, though no one has measured a promised effect.
• Branded prompts surface specific wrong facts you can fix on your own site.

Signs it is not

• Every answer names you. Your prompts are probably too narrow or too branded.
• No answer names anyone useful. Your prompts may be too vague to trigger a shortlist.
• Results change completely every run. Check that you fixed the test conditions.
• The list reads like your keyword spreadsheet. Go back to the buyer's own words.

When you find prompts where rivals are named and you are not, the competitor prompt gap analysis is the next step. Be plain about one limit: a test answer is a sample, not what every buyer sees. Retrieval fires at the engine's discretion, so the same prompt can be answered with or without live search.

Rebuild the full list once a quarter. Fresh calls and tickets bring new problems and new competitor names.

Rerun the frozen core set monthly, or after any of these:

• You change pricing, plans or product names.
• You launch into a new segment or region.
• A major engine changes its default model. Consumer apps change models often, and the model behind "ChatGPT" depends on plan.
• Your site changes how it renders or who it lets in. If a firewall or robots.txt change blocks an AI crawler, answers can shift for reasons that have nothing to do with prompts.

Keep the old list. Comparing this quarter's prompts with last quarter's shows how buyer language is moving. For a quick snapshot between full runs, the free checker on this site shows whether AI names you for up to three buyer questions. It is a point-in-time check, not tracking.