In one sentence: a prompt set is a fixed, versioned list of buyer questions containing no brand names, sized to the quota you can actually afford to rerun, and grouped so that a movement in one theme can be told apart from noise in another.
Everything in AI visibility measurement rests on this list. A dashboard built on a bad prompt set is not slightly wrong; it is wrong in a consistent direction, because the errors that creep in are the ones that flatter you.
Start from decisions, not keywords
The unit is a question a buyer would actually put to an assistant. Those look nothing like keywords: they are sentences, often long, often containing constraints ("under £200", "for a team of five", "that works with Shopify"). In our own Search Console for September 2026, 181 of 475 impression-producing queries were full questions or phrases of seven words or more.
Group them by intent, because the groups behave differently:
| Intent | Example shape | What it tells you |
|---|---|---|
| Category awareness | "What is X and who needs it" | Whether you exist in the model's picture of the category |
| Comparison | "Best X for Y", "A vs B" | Commercial gaps — where you lose to a named competitor |
| Constraint | "Cheapest X that does Y" | Whether your positioning is legible as a set of attributes |
| Problem-first | "How do I fix Z" | Whether you are reachable by people who don't know the category name |
Exclude self-referential questions, at generation time
If the question names your brand, the answer will too. That measures nothing except that the engine can read the question. An LLM asked to generate a prompt set drifts toward these on its own — ours once produced fifteen questions of which thirteen contained the brand name, and the resulting score read 13% across five engines where the honest value was zero.
Filter at generation, not only at scoring. A scoring-time filter still burns the quota on questions you are going to discard, and quota is the binding constraint in this whole exercise.
The fix needs three layers: the generation prompt must be told not to write them, a deterministic filter must catch the ones it writes anyway, and the fallback template set must be clean. Any one layer alone leaks.
Size it to what you can rerun
A prompt set you can only afford to run once is not a measurement, it is an anecdote. Answers vary between runs, so a single sample of a hundred questions tells you less than four runs of twenty-five. Decide the rerun cadence first, then size the set to fit the quota at that cadence.
Version it, and treat changes as breaks
Adding, removing or rewording a question changes what the number means. Version the set, record which version produced each run, and when you change it, start the series again rather than pretending continuity. The temptation to quietly swap in better questions mid-series is strong and it destroys the only thing a trend is good for.
A workable first set
- 20–40 questions, no brand names, mixed across the four intents above.
- Written in the market's language, not translated from English — buyers phrase things differently and translation smooths away exactly the phrasing that matters.
- Stored with an owner and a review date, because a set nobody owns rots into a set nobody trusts.