One prompt tells you almost nothing. Here's what a realistic test actually looks like.
Nobody asks an AI assistant "list all businesses in category X ranked by relevance." Real buyers ask the way they'd ask a knowledgeable friend: "what's the best project management tool for a 5-person team," "can you recommend a good pickleball club in Austin," "who are the top-rated dog groomers near me." These questions are short, conversational, phrased around a specific situation, and — critically — they never mention a specific brand name, because the buyer doesn't know which brand to ask about yet. That's the whole point of the question: discovery, not confirmation.
This matters because it's a fundamentally different test than typing your own brand name into ChatGPT and asking "tell me about Acme Co." That question just tests whether the model knows who you are. It doesn't test whether the model would have named you unprompted — which is the only scenario that actually matters for new customer acquisition.
AI models are non-deterministic by design — the same question asked twice can produce two different answers, sometimes naming a different set of brands, even with nothing else changed. If you test with a single prompt and get a good result, you've learned that the model named you once. If you get a bad result, you've learned it didn't, once. Neither tells you much about what a real distribution of buyers asking a real distribution of questions would actually see.
There's also real variation in how the same underlying intent gets phrased. Someone in a hurry types "best pickleball club austin." Someone more deliberate writes "I'm looking for a pickleball club in Austin — which ones should I consider?" Someone comparison-shopping asks "which pickleball club would you pick and why?" These are the same buying intent wearing different words, and a brand that shows up for one phrasing might not show up for another, depending on how the model retrieves and weighs information for that specific framing.
This is exactly why Mentioned doesn't run a single prompt. When you enter a brand and category, it asks Claude to generate 5 realistic buyer questions for that category — phrased the way a real person searches, never containing your brand name so the test stays organic rather than leading — and sends all 5 to the model independently and in parallel. There's no shared context between them; each is a fresh conversation, exactly like 5 different buyers each asking their own question. Your score is simply how many of the 5 came back mentioning you, out of 5. It's not a weighted estimate or a confidence interval — it's a literal count of what actually happened across a realistic spread of ways a buyer might ask.
If question generation ever fails for some reason, Mentioned falls back to a fixed set of neutral templates ("What are the best {category}?", "Can you recommend a good {category}?", and similar) rather than skipping the check or guessing at a result — see the full breakdown on the methodology page.
The practical takeaway: if you're trying to understand your AI visibility, don't test with one prompt and treat the result as final — in either direction. Ask multiple realistic variations, expect some variance between them, and look at the pattern rather than any single answer. That's a more honest picture of what buyers are actually seeing than a single lucky (or unlucky) prompt could ever give you.