To choose prompts to track for AI visibility, start with the questions real buyers ask at each stage of buying, keep most of them unbranded, begin with 20 to 50, and track your competitors on the same prompts. That set becomes the measuring stick for every AI visibility number you report, and the prompts you choose are the foundation of everything else you measure.

If you've just set up a tracker, or a client wants to know how they show up in ChatGPT and you're staring at an empty prompt list, start here.

Why the prompts you pick decide your AI visibility score

Your AI visibility score is a ratio over the prompts you track, so the prompts decide the score before any AI platform answers a single question.

Two common mistakes show how. Fill the list with branded prompts, such as "is [your brand] good for agencies?", and your score goes up, because an answer to a question that names you almost always mentions you. Fill it with broad awareness questions, such as "what is project management?", and your score goes down, because those answers can name no brands at all.

AI Peekaboo's published methodology makes the same point about its own numbers: scores from different prompt sets shouldn't be treated as directly equivalent without reviewing those differences. The same applies to any tracker. When two trackers disagree about the same brand, a different prompt set is one of the first things to check, and our guide on why your AI visibility tool might be showing you the wrong data covers the other usual causes.

The answers also vary from run to run. A SparkToro study from January 2026 tested this. 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. The study found less than a 1 in 100 chance that ChatGPT or Google's AI, asked the same question 100 times, would give the same list of brands twice. It isn't peer-reviewed, and one of its researchers works at the AI tracking company Gumshoe. Treat your prompt set as a sample of the conversations your buyers have, and read the results as rates across repeated runs. Rand Fishkin, who ran the study, suggests asking at least 60 to 100 times to learn which brands an AI tends to recommend for one prompt. Weekly runs on a single prompt won't get there, so read your results across the whole set, or across a stage or topic over several weeks. Twenty prompts on three platforms give you 60 answers a week, and a stage of six prompts gives 18, or 72 after a month. Treat any single prompt's result as a rough signal.