Designing the prompt set you track
Also known as: query set, tracked prompts
Choosing the questions you measure — the decision that determines whether AI-visibility tracking tells you anything, and the one most often got wrong.
Last revised 2026-09-07. Also available as markdown — request this URL with Accept: text/markdown.
In plain English
Tracking AI visibility means asking the engines a fixed set of questions, repeatedly, and counting who gets named. Everything downstream depends on which questions those are. Ask "what is Acme?" and you learn whether the model has heard of you. Ask "best payroll software for a UK charity" and you learn whether you win the moment a buyer is deciding.
The second question is worth money and the first is worth nothing, yet the first is what most people track, because it is the question they think about all day.
How to fix / set it up
- Write down the last ten questions a real customer asked before buying — from sales calls, not from a keyword tool.
- Rewrite each in the customer's own words, keeping the constraint that makes it specific.
- Sort them by buying stage and check you have at least one from each.
- Freeze the set for a quarter. Record the date you froze it next to the chart.
- Add an objection question about your own brand: it is the one that surfaces hallucinations early.
The technical detail
Write the questions in the buyer's words, not the product's. Buyers describe a problem and a constraint ("for a small team", "under £50", "that integrates with Xero"); they rarely name a category the way a vendor does.
Cover the buying stages: the problem question, the comparison question, the shortlist question, and the objection question ("is X any good", "alternatives to Y"). Visibility usually differs sharply between them, and knowing where you disappear tells you what to write.
Then freeze the set. A chart whose questions changed mid-quarter moves for reasons unrelated to your visibility, and that chart is worse than no chart. Add new questions as additions, never as edits to existing ones.
Fifteen to twenty-five good questions beat a hundred mediocre ones, because every question multiplies by engines and by repetitions: five engines, three samples and twenty-five prompts is 375 answers per cycle. Volume is what tracking actually costs.
FAQ
Should I track my own brand name?
One or two, to catch hallucinations and to see how you are described. Not more — brand-name questions measure recall, not competitiveness, and they flatter you.
How many questions is enough?
Fifteen to twenty-five for one market. Past that you are usually adding variations of the same question, which multiplies cost without adding information.
What if my category has no obvious question?
That is a finding, not an obstacle. If buyers do not ask a question your product answers, the visibility problem is downstream of a positioning problem.