Your keyword tool says "best CRM" gets tens of thousands of searches a month. It has no idea that a sales manager just asked ChatGPT, "What's the best CRM for a ten-person sales team that already lives in Gmail and doesn't have anyone technical to set it up?" That question will never appear in a keyword report. It is also, almost word for word, the kind of question that decides which brands get recommended and which never enter the conversation.
That gap is the whole problem with researching AI search. No AI platform publishes how often each prompt is asked, and the questions buyers type are long, specific and endlessly varied. So the core shift is simple to state: question research is listening, not querying a tool. The best prompt data you will ever find is already inside your business, in sales conversations, support tickets and site searches, waiting to be collected.
This post covers finding and organising buyer questions. Once you have them, our guide to AI share of voice explains how to turn a fixed set into a visibility measurement, and answer engine optimization covers writing pages that answer them well.
Why keyword tools can't see prompts
Keyword research rests on one reliable number: search volume, reported by the search engines themselves. Prompt research, one of the core practices behind the wider AI search optimization picture, has no equivalent. ChatGPT, Perplexity and Gemini do not publish what people ask them, so any "prompt volume" figure you see is a modelled estimate built from panels, sampling and extrapolation. It can be useful for comparing questions within one tool, but a quoted monthly figure could plausibly be half or double the truth.
The shape of demand is different too. Keyword demand concentrates on a few head terms. Prompt demand is spread across thousands of phrasings, each individually small, because every buyer adds their own situation. No single prompt looks important, yet together they represent a large share of buying intent. That is why the method has to change: instead of ranking a list by volume, you gather real questions, find the patterns, and group them.
The anatomy of a buyer prompt
A useful way to think about prompts is as a set of components that buyers combine. A keyword compresses all of this into two or three words. A prompt spells it out, and each component changes which brands an AI assistant is likely to recommend.
| Component | Example | Why it matters |
|---|---|---|
| Who they are | "a ten-person sales team" | Filters to options that fit their size |
| Constraint | "under $30 a user", "no IT staff" | Rules brands in or out |
| Existing setup | "already uses Gmail" | Favours brands known to integrate |
| Goal or problem | "stop leads falling through the cracks" | Links the answer to a use case |
| Comparison target | "cheaper than Salesforce" | Places you beside named rivals |
Breaking prompts into components does two jobs. It shows you which constraints come up most often in your market, which tells you what your content must address explicitly. And it lets you build variations systematically, without guessing, once you know which components real buyers use.
LearnWorlds
Premium all-in-one platform to build, market and sell online courses — with interactive video, a site and mobile-app builder.
Best for: Branded All-in-One Course Academy
Where real buyer questions live
The richest sources are inside your own business, because they capture how people phrase questions before any search box or autocomplete has shaped them.
- Sales conversations. The questions prospects ask on discovery calls are, very often, the same questions they ask AI tools, sometimes word for word. Call notes and recorded transcripts are the single best source for considered, high-value purchases.
- Support tickets and live chat. These reveal the problems customers are trying to solve and the language they use for them, including questions asked before buying.
- On-site search logs. What people type into your own search box shows intent in their exact words.
- Reviews and returns reasons. Your reviews, and competitors', show which attributes buyers care about and how they describe trade-offs.
Outside sources add breadth. Online communities such as Reddit and Quora are full of near-verbatim buyer questions, often including the exact constraints people care about. Long, question-shaped queries in Google Search Console show conversational phrasing that already reaches your site. People Also Ask boxes and the follow-up questions shown under AI Overviews reveal the next step in a buyer's thinking. And direct testing, asking AI assistants variations of a buyer question and noting which follow-ups they suggest, rounds out the picture.
The shift No AI platform publishes prompt volume, so question research is listening, not querying a tool. The best prompt data you will ever find is already inside your business, in sales calls, support tickets and site searches.
Keep observed and invented questions separate
AI assistants are very good at generating plausible buyer questions, and that is both useful and dangerous. Asking a model to suggest variations of a question, or the follow-ups a buyer might ask next, is a fast way to fill gaps and spot phrasings you missed. But a model will just as happily produce dozens of convincing questions that nobody in your market actually asks.
The fix is a simple discipline: keep two labelled lists. Observed prompts are ones you have seen real buyers use, in calls, tickets, search data, reviews or communities. Synthetic prompts are ones a tool generated. Use observed prompts to decide priorities and synthetic ones to fill out the variations within each cluster. Promote a synthetic prompt to the observed list only when you find evidence that real people use similar wording. It sounds pedantic, but it is the difference between building content around demand and building it around a model's imagination.
Organise by stage, then cluster by intent
Raw questions become useful once you group them. Start by sorting them by buying stage: entry questions from people just exploring a problem, comparative questions weighing options, clarifying questions checking details such as pricing, integrations or suitability, and action questions from people close to deciding. Comparative and clarifying questions tend to carry the most commercial weight, because they are where recommendations get made.
Then cluster questions that share an underlying intent, even if the wording differs. "Best CRM for a small team on Gmail" and "which CRM works with Google Workspace for a startup" are the same cluster. The goal is usually somewhere between a dozen and a few dozen clusters, each representing a genuine buyer need, rather than hundreds of individual questions. This is the same topic-first thinking behind a content strategy that compounds, applied to conversational demand.
Decide what to tackle first
Without volume data, prioritising takes judgement. A simple scoring grid keeps that judgement honest and consistent across the team.
Score every cluster from 1 to 3 on each factor, then add them up:
→ Evidence: how often it appears in your own observed sources (1 = once or twice, 3 = constantly)
→ Commercial weight: how close it sits to a buying decision (1 = early curiosity, 3 = choosing between options)
→ Fit: how well your offer genuinely answers it (1 = partly, 3 = exactly what you do best)
→ Gap: how weak your current answer or visibility is (1 = already strong, 3 = absent)
Example: "best CRM for a small team on Gmail" scores Evidence 3 + Weight 3 + Fit 3 + Gap 3 = 12. "What is a CRM?" scores Evidence 2 + Weight 1 + Fit 2 + Gap 1 = 6.
Work from the top down. High-evidence, high-weight clusters where you are currently absent are your priority list. The gap score is where your share-of-voice data plugs in.
Turn clusters into content
Map each priority cluster to a page, or flag it as a gap. One strong page per cluster usually beats a separate thin page for every phrasing, provided the page explicitly addresses the main constraints buyers attach to the question: team size, budget, existing tools, use case. Comparison clusters often call for honest comparison pages, built on the principles in comparison pages that win the buyer search. Product-focused clusters often belong on the product page itself, following product page optimisation and reflecting how AI shopping assistants now surface products.
Use the buyer's own phrasing in headings where it reads naturally, answer each constraint in a self-contained section, and keep the on-page basics in order using the on-page SEO checklist. The citation side of the work, earning mentions on the sources AI tools quote, is covered in the GEO playbook. Questions asked about your brand by name belong to a different job, keeping your brand facts clear and consistent, covered in brand entity optimization.
Keep it fresh
Buyer questions change as markets change: new competitors, new tools to integrate with, new worries. A one-off research project goes stale within months. Refresh your observed list quarterly from the same internal sources, and keep a standing channel where sales and support can drop in new questions as they hear them. That habit is cheap and keeps your content aligned with what people actually ask. Be realistic about outcomes too: much of the value of answering these questions well shows up as influence rather than clicks, the pattern we described in zero-click content.
The short version
Keyword tools can't see the questions buyers ask AI assistants, because no platform publishes prompt volume and any volume figure is a modelled estimate. So question research is listening, not querying a tool. Prompts are long and specific, combining who the buyer is, their constraints, their existing setup, their goal and the alternatives they are weighing, and each of those parts changes which brands get recommended. The best sources are inside your business: sales conversations, support tickets and live chat, on-site search logs, reviews and returns reasons. Communities, long Search Console queries, People Also Ask and AI follow-up questions add breadth. Keep observed and AI-generated prompts in separate lists and prioritise the observed ones. Sort questions by buying stage, cluster them by shared intent, and score each cluster on evidence, commercial weight, fit and current gap. Map priority clusters to strong pages that address the main constraints explicitly, and refresh the research every quarter.
Want content built on what your buyers really ask?
We help brands research buyer questions and turn them into content that gets cited.
Explore Content Marketing →Frequently asked questions
What is prompt research?
Prompt research is the practice of finding the questions people actually ask AI assistants such as ChatGPT, Perplexity and Gemini when they are researching and choosing a product or service. It is the AI-search counterpart to keyword research, but it works differently. Keywords are short, compressed phrases measured by search volume, while prompts are full sentences that carry context, constraints and a request for a recommendation, such as the size of a team, a budget, a tool someone already uses, or a specific problem to solve. Because no AI platform publishes how often each prompt is asked, prompt research relies on listening to real buyer language in sales conversations, support tickets, on-site search, reviews and online communities, then organising those questions into clusters by intent and buying stage. The output feeds two things: the content you create to answer those questions, and the fixed prompt set you use to measure your visibility in AI answers over time.
How are AI prompts different from keywords?
Prompts are longer, more specific and more conversational than keywords, and they usually ask for a judgement rather than a list of pages. A keyword such as 'best CRM' is a few words, while a prompt might be 'what's the best CRM for a ten-person sales team that already uses Gmail and doesn't have an IT department?', often fifteen words or more. That extra detail matters because it contains the constraints that decide which brands get recommended: team size, budget, existing tools, industry and location. Prompts also frequently arrive as part of a multi-step conversation, where later questions refine earlier ones. Demand is spread across thousands of slightly different phrasings rather than concentrated on a few head terms, so no single prompt looks large, but together they represent a great deal of buying intent. Keyword research still matters for classic search; prompt research is a second, complementary half of understanding what buyers want.
Is there search volume data for AI prompts?
Not in the way there is for Google keywords. None of the major AI platforms publishes data on how often particular prompts are asked. Some tools now offer prompt volume estimates, but these are modelled from panels of user behaviour, sampling and extrapolation rather than measured directly, so a figure could easily be several times too high or too low. They can be helpful for relative comparisons within one tool, but they should not be treated as precise numbers or used to justify large decisions on their own. The more reliable signal of demand comes from how often a question appears in your own sources, such as sales calls, support tickets, on-site search logs and customer reviews, combined with how frequently similar questions appear in public communities. Treat any prompt volume figure as an estimate, and prioritise questions you have actually observed buyers asking over ones that simply look plausible.
Can you use AI to generate prompts for research?
Yes, but only to expand coverage, not as evidence of demand. Asking an AI assistant to suggest variations of a buyer question, or to list follow-up questions someone might ask next, is a fast way to fill gaps and spot phrasings you had not considered. The risk is treating those generated prompts as if real buyers ask them. A model will happily produce dozens of plausible questions that nobody in your market actually types, and building content or measurement around them wastes effort. The sensible approach is to keep two labelled lists: observed prompts, which you have seen in sales calls, support tickets, search data, reviews or communities, and synthetic prompts, which a tool generated. Use observed prompts to set priorities and synthetic ones to fill out the variations within each cluster, and promote a synthetic prompt to the observed list only when you find evidence that real buyers use similar wording.
