How to Build a Prompt Set That Actually Measures Your AI Visibility

Start from buyer questions, not keywords

Keyword lists are built for a search box that matches strings. AI engines answer questions in sentences. The prompt "best CRM for small business" is a keyword. The prompt "I run a 12-person agency and need a CRM that handles retainers, what should I look at" is a question a real person asks an assistant. The second one is what you should track, because that is what the engines are actually answering.

Pull real questions from sales calls, support tickets, onboarding forms, and the community threads where your category gets discussed. Write them the way people write them, with context and constraints. Thirty to fifty of these is a good starting range. Fewer and the numbers are noisy. Many more and the weekly run becomes expensive without adding much signal.

Cover the whole funnel

Group prompts by intent so you can see where you are winning and losing.

Category questions: "what are the main options for X." These measure whether you are known at all.

Comparison questions: "X vs Y" and "alternatives to Y." These measure whether you are in the consideration set.

Fit questions: "is X good for a team like mine." These measure whether the engines understand your positioning.

Fact questions: "how much does X cost" and "does X integrate with Z." These measure whether your own pages are the source for facts about you.

A set that is all category questions will tell you about mentions and nothing about citations. A set that is all fact questions will tell you about your pages and nothing about your standing in the market.

Include prompts you expect to lose

It is tempting to only track prompts where you should win. Add some you expect to lose, especially comparison prompts naming a stronger competitor. These are the ones that will move first when the work starts to pay off, and they are the ones that show whether share of answer is actually shifting or whether you are just confirming what you already had.

Keep the wording fixed

Once a prompt is in the set, do not edit it. Engines are sensitive to phrasing, and a small change produces a different answer that looks like a trend and is not. If a prompt turns out to be badly worded, retire it and add a new one, and keep a note of the date. The point of a tracking set is comparability over time, which only holds if the input is stable.

Run each prompt in a fresh session with no history. Run across the engines you care about, in the same mode each time, with search on or off consistently. Record the full answer, not just a yes or no, so you can go back and read why a result changed.

Expect variance and plan for it

The same prompt run twice tends to produce slightly different answers. A company that appears in three of five runs is not "in" or "out," it is at 60 percent. Run each prompt several times per cycle and report the rate. This is the single biggest reason small prompt sets mislead: with ten prompts and one run each, a normal amount of variance looks like a dramatic change.

At AEO GEO Labs a tracking set is typically forty to sixty prompts across four intent groups, run multiple times per week per engine, with results scored for mention, citation, and position in the list. The weekly report shows the rates and the deltas, and the raw answers are kept so any change can be traced to a specific sentence.

Build it, then leave it alone

Write your first forty prompts this week from real buyer language, tag each by intent, and run them once across the engines to get a baseline. Then schedule the run and resist the urge to tweak the wording. The value shows up in week six, when you can see which numbers moved after which changes. If you would rather have the set built and run for you, AEO GEO Labs starts every engagement by building exactly this.