Table of content
We start with topics, not prompts
The common approach in this category is to ask a customer to type in a list of prompts they think matter, then track those. That sounds reasonable until you notice the problem: it asks the customer to already know the answer to the question they’re trying to find out. If you knew exactly which questions AI models were being asked about your category, you probably wouldn’t need a tracking tool in the first place.
We invert that. Instead of starting from a manually written prompt list, we start from a topic (a service, a product category, a use case) and generate the prompts from data. The customer defines what they want to be visible for. We work out how people are actually asking about it.
Where the prompts actually come from
Prompts are generated systematically from keyword and search data tied to the topic, not guessed at. Google Search Console is a key input here: it shows the actual queries a brand’s own site is already getting found for, which is a far more grounded starting point than brainstorming a list of questions that sound plausible. Connecting Search Console is one of the first steps we recommend during onboarding for exactly this reason: the better the underlying data, the better the prompt set.
That data gets translated into the kinds of questions someone would actually type into ChatGPT, Claude, Perplexity, or Google when they’re at that stage of considering the topic, not just keyword variations, but the natural-language questions a real buyer asks.
Why one prompt per topic isn’t enough
A topic isn’t one question. It’s a cluster of related questions asked in different ways by different people, and a prompt set needs to cover two distinct kinds of variation to be reliable.
Synonym-level variation is the same underlying question asked in different words. “Best CRM for a small team” and “which CRM should a small team use” are the same intent, but there’s no guarantee an AI model treats them identically or surfaces the same brands for both. A prompt set that only tests one phrasing is one non-deterministic dice roll away from a misleading result.
Segment-level variation is different. Different types of buyers genuinely ask different questions about the same topic. An enterprise buyer and a solo founder researching the same product category aren’t asking the same question just because it’s nominally about the same thing, they care about different features, different risks, different comparisons. A prompt set built around a single “representative” phrasing collapses that difference and hides how a brand performs for the segments that actually matter to it.
Covering both kinds of variation is why a topic needs a real spread of prompts, not a single best guess. Twenty-five prompts per topic is the minimum where a result starts to reflect a genuine pattern instead of noise from one phrasing that happened to land well, or badly.
Prompts aren’t fixed once and left alone
Because the underlying source data changes, new content gets published, search behavior shifts, a competitor starts ranking for new terms, the prompt set for a topic isn’t a one-time list. It’s re-evaluated as part of the same weekly cycle that re-runs mention tracking, so the questions being tested keep reflecting how people are actually asking about the topic right now, not how they were asking six months ago.
Why this matters more than it sounds like it should
The instinct is to treat prompt selection as a setup step, something you configure once and move past to get to the real dashboard. It’s closer to the opposite. Every score, every competitor comparison, every gap and every recommendation downstream is only as trustworthy as the prompt set it’s built on. A well-designed scoring formula run against a bad, narrow, or guessed-at prompt list still produces a bad, narrow, or misleading number.
Starting from real search data instead of a manually typed list, and covering both how people phrase a question and who’s asking it, is what turns “we tested some prompts in ChatGPT” into a measurement you can actually act on.