How to build a prompt set that reflects real buyers
Your AI visibility data is only as good as the questions you track. Here's how to build a 25-prompt set from sales calls and support tickets, with five archetypes and examples.
The short answer
A good prompt set is 20-30 questions written the way buyers actually phrase them, sourced from sales call recordings and support tickets rather than a keyword tool, and spread across five archetypes: category, comparison, alternatives, problem-led, and jobs-to-be-done. Lock it, version it, and don't swap prompts between cycles or you destroy your own trendline. This is the highest-leverage step in AI visibility measurement and the one almost everyone rushes.
Contents
- Why the prompt set is the whole ballgame
- The five prompt archetypes
- How many prompts do you actually need?
- Where to source prompts
- Sales call recordings (the best source, by a lot)
- Support tickets and docs search
- Your sales team's "what do they always ask?"
- Reddit, community threads, and forums
- Not from a keyword tool
- Rewriting a keyword into a prompt
- Rules for a prompt set you can trust
- What a finished 25-prompt set looks like
- What to do next
The prompt set is the instrument. Everything downstream, mention rate, Share of Voice, competitor rankings, position, is a measurement taken through it. Build it badly and you get beautifully consistent data about questions nobody asks. Most teams spend an afternoon on this and then six months analyzing the output. The ratio should be closer to the reverse.
Why the prompt set is the whole ballgame
Your prompt set decides what "visibility" means for your brand. Track 25 prompts about your strongest use case and you'll look dominant. Track 25 prompts about the comparison and alternatives queries where deals are actually decided and you may find you're invisible. Same brand, same engines, same week, opposite conclusions.
This isn't a hypothetical. We see it in scan data constantly: a brand shows 60%+ mention rate on category questions ("best tools for X") and near zero on comparison questions ("Acme vs Bolt") and alternatives questions ("alternatives to Bolt"). Those last two are where a buyer is 80% of the way to a decision. Being absent there costs revenue in a way that being absent from a top-of-funnel category list does not.
And the stakes are concrete now. In a G2 survey of 1,076 B2B decision-makers, 51% said they now start software research in an AI chatbot, and 33% bought from a vendor they'd never heard of before the AI named it. The prompts those buyers type are the only prompts that matter. Everything else is you measuring yourself.
The five prompt archetypes
A complete prompt set spans five archetypes. Each one catches a different buyer at a different stage, and each one has a different fix when you're absent from it.
| Archetype | What the buyer is doing | Example prompt | If you're absent, the problem is usually |
|---|---|---|---|
| Category | Building a shortlist from zero | "What are the best AI visibility tracking tools for a small marketing team?" | Weak entity presence. The model doesn't associate you with the category at all |
| Comparison | Down to two or three, deciding | "How does Spottlo compare to Profound for tracking ChatGPT mentions?" | No comparison content exists. Nobody, including you, has written the head-to-head |
| Alternatives | Unhappy with an incumbent, or price-shopping | "What are the alternatives to Profound if I want more than one engine?" | You're not mentioned in third-party listicles and roundups, which is where models source these |
| Problem-led | Has a pain, doesn't know the category name yet | "How do I find out if ChatGPT is recommending my competitors instead of me?" | Your content answers product questions, not problem questions |
| Jobs-to-be-done | Has a task, wants the tool that does it | "I need to report AI search visibility to my CMO every month. What should I use?" | No content maps your product to the specific job, with the specific constraint |
Weight them toward the bottom of the funnel. A defensible split for a 25-prompt B2B set:
- Category: 6 prompts
- Comparison: 5 prompts (your top 5 real competitors)
- Alternatives: 4 prompts ("alternatives to X" for the incumbents you displace)
- Problem-led: 6 prompts
- Jobs-to-be-done: 4 prompts
If your category is crowded and price-sensitive, push more into alternatives. If you're creating a category, push into problem-led, because nobody is searching for a category name that doesn't exist yet.
How many prompts do you actually need?
Twenty to thirty per brand. Here's the reasoning, not just the number.
Below ten prompts, one prompt flipping from mentioned to not-mentioned moves your mention rate by 10 points or more. You will spend your Monday explaining a swing that is pure sampling noise. Between 20 and 30, a single flip moves you 3-5 points, which is small enough to be honest and large enough to notice a real change.
Above about 40, you start adding prompts that are paraphrases of prompts you already have. "Best CRM for small teams" and "top CRM tools for small businesses" will return near-identical brand lists. You've doubled your cost and added no information.
The better investment past 25 prompts is depth, not breadth. Running each prompt five times instead of three gives you a much more stable mention rate than adding ten more near-duplicate prompts. Remember that engine outputs are non-deterministic: a single run per prompt is a coin flip, and no amount of prompt breadth fixes that.
Twenty-five is where we landed for Spottlo's per-brand allocation, and it isn't arbitrary. It's enough to cover all five archetypes with real depth in comparison and alternatives, and few enough that a human will actually read every prompt and vouch for it.
Where to source prompts
Sales call recordings (the best source, by a lot)
Go to your call recording tool. Filter to discovery calls that resulted in a closed-won or closed-lost deal. Listen to the first ten minutes, where the prospect describes their situation before your rep has reframed it into your language.
Write down the sentence they use to describe the problem. That sentence, lightly cleaned, is a prompt. "We have no idea if we're showing up when people ask ChatGPT about our space, and my CEO keeps asking" becomes "How do I find out if my company shows up when people ask ChatGPT about our category?"
Pull 30 calls. You'll get 10-15 usable prompts and you'll notice the same three phrasings recurring. Those repeats are your highest-value prompts.
Support tickets and docs search
Your docs search bar is a log of what confused people enough to go looking. Export the top 100 queries. Most are post-purchase and irrelevant, but the pre-purchase ones ("does this work with Gemini", "can I track more than one brand") map straight onto jobs-to-be-done prompts.
Your sales team's "what do they always ask?"
Slack your two best AEs: "What are the three questions every prospect asks you in the first call?" You'll get six sentences back in an hour. All six are prompts.
Reddit, community threads, and forums
Search your category on Reddit and read how people phrase the ask. This has a double benefit: those threads are also a source that AI engines frequently cite, so the phrasing there is closer to what the model has actually seen. The overlap between how buyers ask and what the models were trained on is not a coincidence.
Not from a keyword tool
Keyword tools give you "best crm small business, 4,400/mo". That's a compressed search fragment, not a question. People type full sentences with constraints into a chatbot: "best CRM for a 12-person sales team that already uses HubSpot for marketing and doesn't want to migrate." The constraint clause is the entire difference, and keyword tools have no way to give it to you.
Use keyword data to find topics. Use call recordings to find phrasing.
Rewriting a keyword into a prompt
| SEO keyword | The prompt a real buyer types |
|---|---|
| ai visibility tools | "What tools can tell me if ChatGPT is mentioning my brand?" |
| profound alternatives | "Profound only tracks ChatGPT on their starter plan. What are the alternatives that cover more engines?" |
| geo vs seo | "Do I need a separate strategy for AI search, or is my existing SEO enough?" |
| ai share of voice | "How do I measure how often AI recommends us versus our competitors?" |
| track chatgpt mentions | "My CEO wants a monthly report on whether AI is recommending us. Where do I start?" |
Notice what happens in the right column: length, constraints, and a person behind the question. That's what the model is being asked, so that's what you should be tracking.
Rules for a prompt set you can trust
Lock it. Once the set is live, changing it breaks comparability. If SoV jumps six points the week you added three prompts about your best use case, you've measured your own edit, not the market.
Version it. prompt-set-v1, effective date, list of prompts. When you do change it, bump the version and report both sets side by side for one cycle.
Keep your brand name out of it. With rare exceptions. A prompt containing "Acme" tests whether the model can describe Acme when handed the name. That's not visibility, it's recall, and it will make you feel much better about your position than you should.
Track brand-safety prompts separately. "Is Acme legit", "Acme reviews", "Acme pricing" are worth watching, because the answer can be actively damaging. But keep them in their own bucket, or they'll contaminate your mention rate with prompts you're guaranteed to appear in.
Include prompts you expect to lose. The instinct is to fill the set with questions you'll win. Resist it. The prompts where a competitor owns the answer are the ones that tell you what to build next. A set you score 90% on is a set that has stopped being useful.
Write them as sentences. No keyword fragments. If you wouldn't say it out loud, don't track it.
What a finished 25-prompt set looks like
Here's the shape, for a fictional AI visibility tool:
- Category (6): best AI visibility tools; tools to track brand mentions in ChatGPT; what software tracks Perplexity citations; AI search monitoring tools for agencies; cheapest AI visibility tracker; tools that track Google AI Overviews
- Comparison (5): one per real competitor, phrased as a buyer would: "Is Profound worth $399/mo compared to cheaper AI visibility tools?"
- Alternatives (4): "alternatives to Profound", "alternatives to Semrush's AI visibility toolkit", and the two incumbents you displace most
- Problem-led (6): "why do my competitors show up in ChatGPT and I don't"; "how do I know if AI is sending me traffic"; "my organic traffic dropped and AI Overviews are showing, what do I do"; and so on
- Jobs-to-be-done (4): "I need to report AI visibility to my CMO monthly"; "I manage 8 client brands and need AI rank tracking"; and similar
That's the set. Every prompt is a sentence. Every prompt came from a human. None of them contain the brand name.
What to do next
- Pull 20 discovery call recordings and write down the exact sentence each prospect used to describe their problem. Do this before you open a keyword tool.
- Sort what you have into the five archetypes and find the gap. It's almost always alternatives and comparison, because those feel uncomfortable to track.
- Cut to 25 prompts, kill the near-duplicates, and make sure at least a third are ones you expect to lose.
- Version the file and put a date on it. Treat it like a schema migration, not a working doc.
- Run it across four engines with a free scan and see how the archetypes score differently. The spread between your category score and your alternatives score is your content roadmap for the next quarter.
Frequently asked questions
How many prompts do I need to track? +
Twenty to thirty per brand covers most B2B categories. Below ten, a single answer change swings your numbers by several points and you end up chasing noise. Above forty, you're usually adding near-duplicates that inflate your cost without adding information. Depth per prompt, meaning more runs, beats breadth once you've covered the five archetypes.
Can I use my SEO keywords as AI prompts? +
Not directly. Search keywords are compressed, three-word fragments like 'best crm small business'. People type full sentences into ChatGPT, often with constraints and context. Use your keywords as a starting list, then rewrite each one the way a person would actually ask it out loud, including the qualifier that makes it real.
Where do I find the questions buyers actually ask? +
Sales call recordings are the best source by a distance, specifically the first ten minutes of discovery calls where prospects describe their problem in their own words. After that: support tickets, the search bar on your docs site, Reddit and community threads in your category, and the questions your SDRs get asked on repeat.
Should I include my brand name in the prompts? +
Almost never. A prompt containing your brand name tests whether the model can describe a brand you just named, which is not visibility. The only legitimate use is a small set of brand-safety prompts, like 'is Acme legit' or 'Acme reviews', tracked separately from your visibility set so they don't contaminate the mention rate.
How often should I change my prompt set? +
As rarely as possible. Every swap breaks comparability with prior cycles. Review quarterly, add prompts as additions rather than replacements, and when you do change the set, report both the old and new set side by side for one cycle so the discontinuity is visible instead of mysterious.
Keep reading
How to track brand mentions in ChatGPT
Asking ChatGPT about your brand once tells you nothing. Here's the manual method, why it breaks, and how to build tracking that produces a trendline you can trust.
AI Share of Voice: how to actually calculate it
Most AI Share of Voice numbers are wrong because they count raw mentions. Here's the correct formula, a worked example with real arithmetic, and the four errors that inflate it.
The AI visibility KPIs worth reporting to a CMO
Five metrics that survive contact with an exec, three that get you laughed out of the room, and the one-slide monthly format that makes AI visibility a budget line instead of a curiosity.