The short answer: the best source of questions that customers ask AI assistants is your own customer conversations. Sales call notes, support emails, on-site search, reviews and forum threads contain the exact long, specific questions people type or speak to ChatGPT, Gemini and Perplexity. Collect 30 to 60 of them, group them by buying stage, write answer-first content for each group, and test them in the assistants to see whether you appear.
Traditional keyword research tools were built around short search phrases. People talking to an assistant behave differently. They write whole sentences, add context about their situation and ask follow-ups. “Best accounting software” becomes “I run a three-person design studio in Finland, we invoice in euros and dollars, what bookkeeping tool should we use and what will it cost?” That second question rarely appears in a keyword tool, but your customers say things like it every week.
Why does prompt research differ from keyword research?
Keyword research asks “what do people search for, and how often?” Prompt research asks “what do people want to know before they choose, and in what words?” The difference matters for three reasons:
- Length and context. Questions to assistants contain constraints such as budget, location, company size and urgency. Content that addresses those constraints is more likely to be a match.
- No reliable volume data. Assistants do not publish how often a question is asked. You cannot rank prompts by volume, so you rank them by business value and frequency in your own customer conversations.
- Answers, not links. The goal is to be the source the assistant uses in its answer, which means clear, specific, quotable content.
Source 1: What do your sales calls and enquiries reveal?
This is the richest source. Go through the last 20 to 30 sales calls, discovery meetings or enquiry emails and list every question a prospect asked before buying. Capture it in their words, not yours. Typical patterns:
- “How much does this usually cost for a company our size?”
- “What is the difference between you and doing it ourselves?”
- “How long until we see results?”
- “What do I need to prepare before we start?”
- “What happens if it does not work?”
Every one of these is a question an assistant will be asked too. If your answers live only in your head and in sales calls, no assistant can quote them.
Source 2: What does support and onboarding teach you?
Support emails, chat transcripts and onboarding questions show what people struggle with after they decide. They also show terminology: the words customers use for your product, which may differ from your marketing language. If customers call your “retainer” a “monthly plan”, an assistant will see “monthly plan” in many questions and may not map your page to it.
If you do not have a support team, use your inbox. Search for words like “how”, “can you”, “what if” and “is it possible” over the last six months.
Source 3: What do on-site search and Search Console show?
If your website has a search box, the queries typed into it are direct statements of need. In Google Search Console, filter the Performance report for queries containing question words (how, what, why, can, should, does) or with more than five words. These long queries are the closest thing to prompts in your existing data.
They also show what you already rank for, which tells you where you already have some authority to build from.
Source 4: What are people saying in reviews and forums?
Read reviews of your business and your competitors’ businesses. Look for the sentence that starts “I wish I had known…” or “the thing nobody told me was…”. Those are unanswered questions.
Forum and community threads, including Reddit, are particularly useful because answer engines frequently draw on them. We cover that in our piece on why AI answer engines cite Reddit so much. Search for your product category and note the recurring questions, the objections and the comparisons people make. Do not quote individuals or copy their wording into your content; use the themes.
Source 5: What do the assistants suggest themselves?
Ask ChatGPT, Gemini or Perplexity: “What questions do people usually ask before choosing a [your service]?” and “What should I check before hiring a [your service provider]?” The answers will not be precise data, but they reveal how the assistants frame your category. Treat the output as a prompt for your own thinking, not as evidence of demand, and verify against your real customer conversations.
How do you organise the questions?
Put everything into one spreadsheet with four columns: the question in customer wording, the source, the buying stage and the page that answers it (or “none”). Group by stage:
| Stage | Typical question | Content type |
|---|---|---|
| Learning | What is X and how does it work? | Explainer |
| Comparing | X versus Y, which is better for my situation? | Comparison, decision guide |
| Pricing | How much does X cost? | Cost breakdown |
| Choosing a provider | How do I pick a good one, what are red flags? | Checklist |
| Doubt | Is X worth it, what if it fails? | Honest assessment |
The stages nearest to purchase, pricing, comparing and choosing, deserve priority. They are also the stages where businesses most often say nothing, because they feel uncomfortable discussing cost or competitors.
How do you test whether you show up?
Take your top 20 questions and ask each in two or three assistants. Record for each: is your brand mentioned, is your site cited as a source, are competitors named, and is anything said about you that is wrong? Do it in a fresh session each time, because earlier conversation influences answers. Responses vary from run to run, so treat single results as snapshots. We describe a repeatable routine in our 45-minute monthly AI visibility check.
If an assistant says something incorrect about you, our correction playbook covers what to do.
How do you turn the list into content?
For each cluster of similar questions, write one page that answers the core question directly in the first two paragraphs, then covers the variations under their own headings, ideally phrased as questions. Include specifics: numbers, ranges, conditions and examples. Vague pages are rarely quoted. Our post on how AI answer engines read your pricing page shows how this applies to the most sensitive topic of all.
How big should the list be, and how often should you refresh it?
A working list of 30 to 60 questions is enough for a small business. More than that becomes a spreadsheet nobody uses. Review it every quarter: add the new questions from recent calls, remove those you have answered well, and re-test the top ones. Questions shift when competitors launch products, when platforms change, and when the news cycle makes people curious about something new.
The practical advantage of this method is that it costs almost nothing but attention. The questions are already being asked in your business every week. Writing them down is the first step toward being the answer.
A worked example of turning questions into pages
Suppose a bookkeeping firm for freelancers lists these questions from sales calls: “Do I need an accountant if I only invoice a few clients?”, “How much does bookkeeping cost for a sole trader?”, “What records must I keep, and for how long?” and “Can I switch accountants mid-year?” The first and fourth sit in the comparing and doubt stages; the second is pure pricing; the third is learning. The firm writes four pages, each opening with a direct answer in the first two paragraphs, a short table of typical cost ranges clearly labelled as ranges, and a plain statement of what changes the answer. Within a quarter they can test the same four questions in the assistants and see whether they are mentioned. The method is deliberately modest: it works because it starts from real questions and ends with something checkable, not because it uses clever tooling.