How to build an AI visibility prompt set from real buyer language

A good AI visibility prompt set is a sample of the questions your buyers actually ask, in the words they actually use. Most sets are not. They are written by marketers, in category jargon, often generated from a keyword list, and they measure how the brand looks to people who already know what to search for.

The method below fixes that. It is the approach the more rigorous practitioners converge on, and it is what the RateMyPrompts rubric is built to reward. None of it needs a paid data source.

Why does the wording of a prompt matter so much?

Because the assistant answers the question it is given, and different phrasings pull different sources. A prompt written in category language, such as "best CRM software for small agencies 2026", tends to retrieve list articles, which are dense with brand names. A buyer describing their situation, such as "we're a 12-person agency still running everything off spreadsheets, what should we move to", tends to retrieve explainers, which name far fewer brands.

So a jargon-heavy set returns a flattering visibility number that does not describe what your buyers see. We learned this the hard way: an earlier version of our own rubric scored the jargon version of ten buyer needs higher than the natural version, 56 against 52, because it rewarded prompts containing the category noun. We changed the rubric. In one live 156-prompt set we reviewed, only 13 of 94 unbranded prompts contained the category word at all; the rest described situations.

The eight steps

1. Start from the decision, not the prompt

Write down what you need to learn before writing a single prompt: which markets, which buyer roles, which parts of the journey, and which competitors matter. The set is a measurement instrument for those questions, so they come first.

2. Mine your own buyer language

The best prompts are already in your systems. Go through:

3. Add observed demand from outside

Next, bring in questions people ask where you can see them: long-tail queries from the Search Console performance report rewritten as the question behind them, People Also Ask boxes, and threads on Reddit or industry forums. Use them as evidence of what people ask, never pasted in raw.

4. Keep the buyer's words

This is the step most teams skip. If a buyer asks for an outcome without naming your category, do not rewrite the prompt to include the category term. That single edit turns a buyer's question into a marketer's one, and it moves the number.

5. Add real constraints

Real questions carry conditions: company size, budget, role, industry, the tools already in place, location and compliance needs. "Accounting software for a 40-person charity that needs Gift Aid claims" is a real question; "best accounting software" is a keyword. The Prompt Fit Score rewards sets where a good share of prompts are specific in this way, because those are the questions where recommendations differ.

6. Cover the whole journey

Spread the set across the stages buyers pass through: discovery, problem, evaluation, comparison, validation and post-purchase. Published guides disagree on the right split, so do not chase a ratio. Occupy at least four of the six stages and keep any one stage under about 45% of the set. Most teams over-invest in comparison prompts and ignore the rest.

7. Keep what recurs, not what is popular

There is no auditable search-volume figure for AI prompts, and you do not need one. A question is worth tracking when it turns up in more than one independent source: in calls and in tickets, or in site search and on forums. Use AI generation only to fill gaps the evidence leaves, and label those prompts so you can review them later.

8. Size it, then freeze it

Aim for at least 30 prompts, 50 or more if you will compare topics or markets, and five or more per topic you plan to report. This matches decades of information retrieval research, which found a good test needs at least 25 queries, and 50 is better. Then freeze 70 to 80% of the set word for word as a stable core, and treat the rest as a discovery slice you change deliberately.

A worked example

A support ticket reads: "We moved from spreadsheets last year and now the team keeps double-booking the meeting rooms. Is there something that works with Google Calendar for about 30 people?"

The prompt keeps the buyer's framing and the constraints:

What room booking tool works with Google Calendar for an office of about 30 people who keep double-booking rooms?

It does not become "best meeting room booking software 2026". It names a situation, a size and an integration, and it sits at the problem stage. Ten prompts like this, drawn from real tickets and calls, will tell you more about your visibility than a hundred generated from a keyword list.

What to do this week

1. List the decisions the set must inform: markets, roles, stages, competitors.
2. Pull 20 recent call transcripts and 50 recent tickets, and copy out every question in the buyer's own words.
3. Add questions from Search Console and forums that recur in your first-party sources.
4. Check each prompt for a constraint, and keep your own brand name out of most of them. Branded prompts are covered in why branded prompts inflate your AI visibility score.
5. Tag each prompt with a topic and a journey stage, and check no stage dominates.
6. Freeze the core and date the version.

When you have a draft, run it through the free RateMyPrompts scorecard. It will tell you how buyer-shaped the set is, which prompts are near-duplicates, and where the journey has gaps, with the points each fix is worth.