Traditional keyword research starts with a search volume tool: type in a seed term, get a list of related queries ranked by monthly searches. That process assumes people type short, fragmented phrases into a search box. AI search breaks that assumption. People ask ChatGPT and Perplexity full sentences, with context, constraints, and follow-up questions. AI keyword research has to start somewhere else entirely.
Why keyword tools miss the target
A Google keyword tool tells you "best CRM software" gets 8,000 monthly searches. It tells you nothing about "what CRM should a 12-person real estate team with a tight budget use that integrates with our existing email tool," which is closer to how someone actually phrases a question to an AI assistant. The volume data that drove SEO for two decades doesn't capture the long, conversational, context-rich queries that dominate AI search.
Where to actually find the prompts
Sales call transcripts and support tickets. The exact phrasing your prospects and customers use to describe their problem is the single best source of real prompt language. If your sales team hears "how do I get my team to actually use this without a ton of training," that's a prompt, not just a talking point.
Competitor comparison pages, read for structure, not content. If competitors have built pages comparing themselves to five alternatives, the questions those pages answer are the questions buyers are asking. Extract the question, not the competitor's spin on the answer.
Onboarding and demo objections. The reasons prospects hesitate before buying are almost always phrased as questions somewhere in your sales process. "Is this secure enough for a regulated industry" is a prompt. Write it down every time you hear a version of it.
Reddit, forums, and review site Q&A sections. These are the closest public proxy to how people phrase requests to AI assistants: informal, specific, and full of real constraints ("budget under $50," "works with existing stack," "for a small team").
Direct testing across engines. Ask ChatGPT, Claude, Gemini, and Perplexity broad category questions and read the follow-up questions they suggest or the sub-topics they cover in their answer. These reveal how the models themselves decompose a broad topic into the specific sub-queries worth targeting.
Structuring the prompt set once you have it
Group the prompts you've gathered into four buckets, because each requires different content:
- Category-definition prompts ("what is X"): these need a clear, quotable definition early in a dedicated page.
- Comparison prompts ("X vs Y", "alternatives to X"): these need explicit, factual comparison content, not vague positioning.
- Recommendation prompts ("best X for Y"): these need proof, specificity about the "for Y" qualifier, and enough detail to be trusted over a competitor.
- Problem-first prompts ("how do I solve Z"): these need to lead with the solution before introducing the product, because the person hasn't decided they need a product yet.
Validating the list before you build content
Before writing anything, run each candidate prompt through the major AI engines and record whether you already appear, whether a competitor appears instead, and what source gets cited. This turns your keyword list into a prioritized gap list: the prompts where you're currently invisible and a competitor is present are your highest-value targets, because the demand is proven and the opportunity is measurable.
The discipline that matters most
AI keyword research isn't a one-time list. Buyer language shifts as your category matures, as competitors publish new comparison content, and as AI engines change what they surface. Revisit your prompt set quarterly, retire prompts that no longer convert attention into anything useful, and keep pulling fresh phrasing from the same sources: real conversations with real prospects, not a volume estimate from a tool built for a search box that fewer and fewer of your buyers are typing into first.