Measurement

AI prompt research: how to find the questions buyers actually ask AI

By Arnav Mukherjee, founder of TofuBofu · July 13, 2026

Early in building TofuBofu, I ran our target queries through Google Keyword Planner to size the market, the way I would have for any SEO project. It reported zero volume. Not low volume: zero, for question after question that I knew real buyers were asking, because I was watching them ask AI engines those exact questions. I wrote up why Keyword Planner gives you the wrong answer for AI search. This post is the other half: what to do instead.

Keyword research answers "what do people type into a search box." Prompt research answers "what do buyers tell an AI about their situation." Those are different questions with different raw material, and the second one has no Keyword Planner. The good news: for a services firm, the best prompt data does not live in a tool at all. It lives in places you already have access to.

A prompt is not a long keyword

The structural difference matters more than the length. A keyword compresses intent down to its category. A prompt carries the whole situation: who the buyer is, what just happened, what they are constrained by, and what they want. The engine uses every part of that to decide who gets recommended, which means every part of it is something your content can match, or miss.

Read the diagram bottom row again, because it is the whole strategy in miniature. The context is won by vertical pages. The constraint is won by proof. The ask is won by being a name engines trust enough to put on a shortlist. Prompt research is how you find out which contexts, triggers, and constraints your buyers actually bring, so you build those pages and not generic ones.

The five sources, in order of trust

1. Your own sales conversations

Discovery calls, inbound emails, the 'we found you because' answers. This is ground truth: the words buyers use before they have adopted your vocabulary. Pull the last twenty and write down the situation each buyer described in their own phrasing. Those situations are prompts.

2. Reddit and community threads

Search Reddit for recommendation requests in your category ('recommend an MSP', 'accountant for SaaS startup'). The posts are literally prompts, context and constraints included, written in public. As a bonus, these same threads are among the sources AI engines cite most, so you are reading the engine's own source material.

3. The engines themselves

Ask ChatGPT or Claude: 'What questions do operations leaders at mid-size law firms ask when choosing an IT provider?' The engines have seen millions of these conversations and produce realistic phrasings, including constraint patterns you have not thought of. Treat the output as hypotheses to verify against sources 1 and 2, not as data.

4. Your search-console queries, reframed

Google Search Console shows the keyword form of your demand. Reframe each meaningful query as the situation behind it: 'msp for dental office' becomes a dental practice manager describing a compliance-heavy migration. The keyword tells you the category exists; the reframe gives you the prompt.

5. Prompt-volume datasets, for direction only

SEO platforms like Semrush now sell prompt databases built from clickstream panels, hundreds of millions of prompts sampled from real usage. Useful for spotting big patterns and category language. But the panels are samples, the coverage is young, and niche B2B situations are exactly what sampling misses. Never let a zero in a prompt database talk you out of a question your own buyers demonstrably ask.

See how you score on real buying prompts, free

TofuBofu researches your company, generates the buying-intent prompts for your category, and runs them across six AI engines. You can edit every prompt before the scan.

Run your free scan

Building the set: 10 to 25 prompts, structured by funnel

Resist the urge to track hundreds. Because AI answers vary between runs, a reliable read means sampling each prompt several times per engine, and that math punishes bloated sets. Twenty prompts sampled properly beat a thousand sampled once. Structure the set the way your funnel is structured:

Category head prompts ("best managed IT providers in Ontario") tell you whether you exist in the engine's mental map at all. Situation prompts (the law-firm example above) are where deals actually are, and where specific firms beat big brands. Constraint prompts ("MSP that handles HIPAA compliance") test whether your proof is legible. Competitive prompts ("alternatives to [your biggest competitor]") test whether you win the switchers. Weight your attention toward the bottom of that list, the same buying-intent weighting our scoring uses, because those are the prompts closest to a signed contract.

Then keep the set stable. The value compounds when the same prompts are asked month after month, because only then does a change in the answers mean something changed in the world, rather than in your questionnaire. Add prompts when your business genuinely changes; do not churn them for novelty. The full measurement mechanics, sampling, per-engine scoring, cadence, are in how to measure AI visibility.

Frequently asked questions

What is AI prompt research?

The AI-search equivalent of keyword research: figuring out what your buyers actually ask AI engines when they research a purchase, so your content and third-party presence can answer those exact questions. The difference is the raw material. Keywords are two to five typed words; prompts are full sentences carrying context, constraints, and situation, and no public tool reports their volume the way Keyword Planner reports search volume.

How are AI prompts different from search keywords?

Prompts are longer, conversational, and loaded with context. A Googler types 'it support toronto'. The same buyer tells ChatGPT they run a 40-person law firm in Toronto, their IT person just quit, and they need someone who understands legal-industry compliance. The prompt names the industry, the company size, the trigger event, and the constraint, and the engine uses all of it to pick who to recommend. Content that matches only the keyword misses everything the prompt actually asked.

Is there a Keyword Planner for AI prompts?

Not a definitive one. Google Keyword Planner reports zero for most conversational queries, which makes it a false-negative machine for AI search. SEO platforms like Semrush have started building prompt datasets from clickstream panels, useful for direction but young, sampled, and nowhere near the coverage keyword databases have for search. For a services firm, direct sources, your own sales conversations, communities, and the engines themselves, are more accurate than any volume database today.

How many prompts should I track?

For a focused services firm, 10 to 25 well-chosen prompts cover the territory that matters: your category head questions, your highest-intent buying situations, your key verticals and locations, and alternatives-to questions around your competitors. Depth beats breadth, because each prompt should be sampled several times per engine to handle answer variability. A thousand prompts sampled once tell you less than twenty prompts sampled properly.

Where do I find the prompts buyers actually use?

Five reliable sources: your own sales calls and inbound emails (the phrasing buyers use before they found you), Reddit and community threads where people ask for recommendations in your category, the questions your buyers ask you in discovery meetings, the AI engines themselves (ask them what questions companies like your buyers ask about your category), and your existing search-console queries reframed as situations. Each source produces phrasings a keyword tool never shows.

Do I need different prompts for each AI engine?

No. Buyers ask the same kinds of questions everywhere, so one well-built prompt set works across engines. What differs per engine is the answer, which is why you run the same set on all of them and score each engine separately. Keeping the prompt set constant is what makes month-over-month comparison meaningful.

Sources and further reading

Semrush: How to do prompt research for AI SEO: the tool-vendor view, including their clickstream-based prompt datasets
G2 2026 AI Search Insight Report: 51% of B2B buyers now start research with an AI chatbot, the demand prompt research maps
Forrester: The State of Business Buying, 2026: 94% of B2B buyers use AI somewhere in the buying process