When you plan AI visibility tracking, you need the questions you monitor to represent decisions your buyers actually face. Choose prompts that reflect your buyers' questions. The method below helps you select those questions and calculate how many you can track.
Start from decisions, not keywords
A keyword is what someone types into a search box. A prompt is a question someone asks an assistant, and the two do not look alike.
People bring assistants the messy version of a decision. Not "project management software" but "we're a 12 person agency drowning in client work, what should we use to keep track of it". The extra context is the point, because it is what the answer responds to.
So build the set from decisions your buyers face. Four sources are reliable:
- Sales calls. The question a prospect asked in the first ten minutes, in their words.
- Support requests from people who chose wrongly, which reveal the comparison they got wrong.
- The comparison your buyers make. If they always weigh you against two named alternatives, that comparison is a prompt.
- The objection you answer weekly. Phrased as the buyer's question, not as your rebuttal.
Cover the four question types
A set drawn from one type gives a distorted picture. Aim for coverage across these:
- Category questions. "What tools do X teams use for Y?" These have the widest audience and are the hardest to appear in.
- Comparison questions. "Is A or B better for Z?" Fewer people ask them and those who do are close to deciding.
- Problem questions that do not name a category at all. "How do I stop losing track of client work?" Your buyer may not know the category exists.
- Brand questions. "Is [your company] any good?" and "[your company] vs [competitor]". These matter most, because the answer is what a buyer sees at the point of decision, and they are the ones most teams forget to track.
Brand questions are also where the fastest wins are. An answer describing your product inaccurately or out of date is a concrete problem with a traceable source.
Add the persona where it changes the answer
Assistants answer conditionally. The same category question asked as an enterprise procurement lead and as a solo consultant returns different sets of names, and both answers are correct for their conditions.
Include the persona in the prompt wording when your market divides that way. If small and large teams need different recommendations, track prompts for both. A prompt without those conditions does not distinguish their needs.
Do not multiply personas beyond the ones that change the answer. Every persona variant multiplies the collection cost.
How many you can afford, with the arithmetic
Cost is a multiplication: prompts × providers × days per week.
Worked, against a 10,000 credit weekly allowance:
- 100 prompts, three providers, five days: 1,500 a week. Comfortable, and leaves most of the allowance for keywords.
- 500 prompts, three providers, five days: 7,500 a week. Most of the allowance for prompts alone.
- 500 prompts, three providers, twice a week: 3,000 a week. The same coverage, checked often enough for a trend.
The comparison between the last two is the decision most teams get wrong. Reducing collection from five days to two keeps the same prompt set and costs 60 percent less. It also produces fewer observations; choose the frequency your reporting needs rather than assuming the two schedules provide equivalent evidence.
Start narrow. Thirty prompts you chose deliberately beat three hundred generated from a keyword list, and you can always add.
Write them so the series stays comparable
Three rules, all following from the fact that the prompt is the measurement instrument:
Fix the wording. Changing a prompt restarts its series, because a different question is now being asked. Record the date whenever you deliberately change one.
Write them as a person would ask. Keyword-shaped prompts get keyword-shaped answers that do not resemble what your buyers see.
Avoid leading questions. "Why is [your product] the best option for agencies?" will produce a flattering answer that measures nothing. Ask the question your buyer would ask, including the ones whose honest answer is not in your favor.
Review the set quarterly
Drop prompts where the answer has not changed in months and nothing you could do would change it. Add the questions that started appearing in sales calls. Keep the core set stable, because the long series is what carries the evidence.
Review the prompt set before committing recurring usage. Each question should serve a defined buyer decision and remain useful when you compare its answers over time.
See how persona context changes the question, why repeated answers vary, or AI rank tracking.