Keyword research tells you what people type into a search box. It does not tell you what they ask an AI assistant. Prompts are longer, more conversational and full of context: “I run a three-chair dental practice in Leeds and my website gets no bookings — should I hire an agency or try Google Ads myself?” No keyword tool shows that exact question, and no AI company publishes its prompt logs.
That does not mean prompt research is guesswork. The questions people ask AI are usually the same questions they ask your sales team, your support desk and online communities. The job is to collect those questions systematically, group them, and pick the ones that matter for your business.
This guide covers the practical sources of prompt ideas, how to turn them into a tracked prompt set, how to prioritise, and what limitations to keep in mind.
Key Takeaways
- No AI platform publishes prompt volumes, so prompt research combines first-party data, search data and AI tool outputs.
- Sales calls, support tickets and reviews are the richest sources of real buyer language.
- Search Console’s long, question-style queries and People Also Ask show how people phrase problems in search.
- Bing Webmaster Tools’ AI Performance report shows sampled grounding queries that led to citations of your site.
- Prioritise prompts by business value, intent and whether you can credibly answer them — then track a small, stable set.
What prompt research is (and isn’t)
Prompt research is the process of finding the questions your audience asks AI assistants and turning them into a set you can create content for and track. It is not a replacement for keyword research — search demand still matters, and Google says its AI features rely on the same SEO best practices as the rest of Search. Instead, it adds a layer: understanding the context, constraints and comparisons people include when they talk to an assistant.
Two points shape the method. First, assistants break prompts into sub-searches: Google describes AI Mode and AI Overviews “issuing multiple related searches across subtopics and data sources”, and OpenAI says ChatGPT search “typically rewrites your query into one or more targeted queries” when it uses search providers. So you need to cover the underlying sub-questions, not just the headline prompt — see query fan-out content. Second, there is no official prompt-volume data, so any “AI search volume” figure from a tool is an estimate.
Where to find real prompt ideas
| Source | What it gives you | How to collect it |
|---|---|---|
| Sales calls and demos | Buyer questions, objections, comparisons | Tag questions in call recordings or CRM notes; export monthly |
| Support tickets and live chat | How-to and troubleshooting questions in customers’ words | Export ticket subjects; group by topic |
| Reviews (yours and competitors’) | Decision criteria and pain points | Read recent reviews; note recurring phrases |
| Search Console | Long, question-style queries that already reach your site | Filter queries by length or question words with regex |
| People Also Ask and related searches | Follow-up questions Google associates with a topic | Record PAA questions for core topics |
| Communities and forums | Unfiltered questions and context | Search Reddit, Quora and niche forums for your category |
| Bing AI Performance report | Sampled grounding queries behind citations of your site | Review in Bing Webmaster Tools |
| AI tools themselves | Follow-up suggestions and the sub-questions they research | Run seed prompts and note suggested follow-ups |
Search Console: the long-query goldmine
Google Search Console’s Performance report shows real queries, and conversational ones are increasingly common. Use a regex filter to surface questions — for example, queries beginning with how, what, why, which, can, should or best — or queries with many words. Our guide to Search Console regex filters has ready-made patterns. Note that Google’s newer Generative AI performance report shows AI Overviews and AI Mode impressions by page, country and device — it does not list prompts or queries.
Bing: grounding queries
Microsoft’s AI Performance report in Bing Webmaster Tools, launched in public preview on 10 February 2026, includes “grounding queries”: the key phrases the AI used to retrieve your content, based on a sample of citation activity across Copilot, Bing AI summaries and some partner integrations. It is one of the few first-party views of how an AI system searched for your pages.
People Also Ask and question keywords
People Also Ask boxes and question keywords show how Google connects follow-up questions. They make good sub-questions for each priority prompt.
Step-by-step: build your prompt set
- Collect raw questions. Pull 100–300 questions from the sources above. Keep the original wording.
- Add context. Rewrite each into a realistic prompt with the details users typically add — role, budget, location, constraints.
- Group by topic and intent. Cluster into informational, comparison, recommendation and how-to prompts. Our guide to category entry points helps map these to buying situations.
- Test in AI tools. Run each candidate prompt in ChatGPT, Google AI Mode, Gemini, Perplexity and Copilot. Record whether you are mentioned or cited, who is, and which sources appear.
- Score and prioritise. Use the scoring table below.
- Choose a tracked set. Keep it small enough to review regularly.
- Map prompts to content. For each priority prompt, identify the page that should answer it and the sub-questions it must cover.
- Re-run monthly. AI answers vary between runs, so track trends over several checks rather than single results — see AI answer variability.
How to prioritise prompts
| Criterion | Question to ask | Score 1–3 |
|---|---|---|
| Business value | Would winning this prompt lead to revenue or qualified leads? | 3 = direct buying prompt |
| Intent stage | Is the user comparing or choosing, not just learning? | 3 = comparison or recommendation |
| Right to win | Do we genuinely have the best answer or product for this? | 3 = clear strength |
| Current gap | Are competitors named and we are not? | 3 = we are absent |
| Evidence available | Do we have data, case studies or expertise to support content? | 3 = strong proof |
Semrush makes a similar point in its August 2026 guide to prompt research, recommending a focus on bottom-of-funnel comparison and “best” prompts and reporting that, in its internal testing, “10 well-chosen prompts per product are usually enough” to see whether AI consistently recommends a brand. Treat that as one vendor’s rule of thumb, not a standard.
“Be wary of third-party tools that promise ranking success or claim to use "internal" Google metrics.”
— Google Search Central, Google's guide to optimizing for generative AI features on Google Search
Prompt tracking tools can save time, but their volumes and visibility scores are modelled estimates. Use them alongside your own first-party sources.
Turning prompts into content
- Answer the main prompt directly near the top of the relevant page.
- Cover the sub-questions AI is likely to research — costs, comparisons, steps, risks.
- Use the language customers use, including constraints such as budget or team size; see conversational query content.
- Add FAQs drawn from real questions — see FAQ content in the AI era.
- Track mentions and citations for your tracked set — see AI share of voice.
Common mistakes
- Inventing prompts in a meeting. Start from real customer language, not internal jargon.
- Tracking hundreds of prompts. A huge set gets ignored; keep a focused core.
- Treating tool volumes as fact. No platform publishes prompt volumes.
- Judging from a single run. Answers change between sessions and users.
- Only tracking branded prompts. Buyers often ask category questions before they know your name.
- Forgetting search. Keyword research still matters; see our keyword research guide.
Related Guides
- Help Centers and Documentation: Content AI Assistants Rely On
- FAQ Content in the AI Era: What Still Works After Rich Result Changes
- Press Releases and AI Search: Do Newswires Still Help?
Frequently Asked Questions
Is there a tool that shows exactly what people ask ChatGPT?
No. OpenAI and other AI companies do not publish prompt data. Tools that show prompt volumes are estimating, so combine them with first-party sources such as sales calls and support tickets.
How is prompt research different from keyword research?
Keyword research measures search demand for short phrases. Prompt research captures longer, conversational questions with context and constraints. You need both.
Does Google Search Console show AI Mode prompts?
Google’s Generative AI performance report shows impressions by page, country and device, not queries. The main Performance report still shows search queries, including long conversational ones.
How many prompts should I track?
Start small — a core set per product or service that you can review monthly. Semrush suggests around 10 well-chosen prompts per product based on its internal testing.
How often should I refresh my prompt set?
Review it quarterly and whenever you launch a product, enter a new market or see new questions in sales and support.
Start with the questions you already hear
Your customers have been telling you their prompts for years — in calls, tickets and reviews. Collect them, test them in AI tools, prioritise by value and track a focused set. For the measurement side, see using Search Console data for AI search. If you want help building a prompt set and content plan, request a custom quote.
References
- AI features and your website — Google Search Central
- ChatGPT search — OpenAI Help Center
- Generative AI performance report (Search) — Search Console Help
- Introducing AI Performance in Bing Webmaster Tools (public preview, 10 February 2026) — Bing Webmaster Blog
- How to do prompt research for AI SEO — Semrush blog (17 August 2026)
- Google's guide to optimizing for generative AI features on Google Search
