Keyword research tells you what people type into a search box. Prompt research tells you what they ask an assistant, which is usually longer, more specific and closer to a decision.
Start from real questions
- Sales call notes and recordings
- Support tickets and live chat logs
- Site search queries
- Reddit threads, forums and review sites in your category
- Search Console queries of six words or more
These sources give you the language buyers actually use, including the constraints they add: budget, location, team size, integrations.
Expand with fan-out thinking
For each core question, write down the sub-questions an AI system would likely search to answer it. A prompt like "best payroll software for a 50-person UK company" fans out into pricing, HMRC compliance, integrations, reviews, alternatives and setup time.
Each sub-question is a retrieval opportunity. If you have no page that answers it, you cannot be the source for that part of the answer. My article on query fan-out covers this in more depth.
Build a prompt set
Group prompts by topic and stage: problem-aware, comparing options, choosing a provider. Aim for 20 to 50 prompts per priority topic. Keep the wording natural. Record which prompts carry commercial intent.
Common mistakes
- Inventing prompts in a meeting rather than taking them from customers.
- Only tracking branded prompts.
- Treating one prompt as a keyword with a fixed ranking. Answers vary between runs.