Definition
Prompt sampling is a measurement method for AI visibility in which you ask a fixed set of questions (prompts) repeatedly to several AI engines and record the answers systematically: which brands are named, which sources are cited, in what order and in what context. Because a language model gives a slightly different answer each time it is asked the same question, one answer is not a measurement; a sample of answers is.
The method is necessary because classic ranking data does not exist for AI answers. In Google a page is at position 3 or it is not; in ChatGPT a brand is named in 60% of answers, in Perplexity in 20% and in Gemini never. Only by asking the same prompt dozens of times, spread over days and engines, do you get a reliable percentage instead of a lucky hit.
A good setup has four parts. A prompt set that reflects your customers' real questions, per stage of the customer journey (orientation, comparison, choice) and per topic — usually 50 to 500 prompts. A list of engines and modes (ChatGPT with and without search, Gemini, Perplexity, Google AI Mode and AI Overviews, Claude, Copilot). A sample size per prompt, for example five to ten repetitions per week. And a fixed classification of the answers: mentioned or not, cited or not, position in the answer, sentiment and which competitors appear alongside.
From the sample you calculate metrics such as mention rate (percentage of answers in which your brand appears), citation rate (percentage in which your domain is a source), share of voice (your mentions as a share of all brand mentions) and average position. Trends over time matter more than absolute values, because models and their search behaviour change monthly.
There are limitations too: answers differ per country, language, logged-in user and device, and API costs grow with large sets. Traze measures AI visibility with prompt sampling per engine and per topic, ties every citation back to the page that earned it, and uses the outcome to decide which content needs to be written or rewritten.