Track the prompts that hold still

It is the practice of running a fixed set of questions against AI engines on a schedule and recording how your brand appears in the answers. TrustData runs each prompt in two modes, with live web search and without, across four question types, and keeps the full response behind every result. Prompt counts run from 50 on Measure to 400 on Prove, against monthly probe budgets of 3,000 to 24,000.

It is the median pairwise similarity between the brand lists an engine returned across a prompt's recent probes. If an engine returns roughly the same set of brands each time, the prompt is stable and a trend drawn from it means something. If the set changes every run, the prompt is unstable and its chart is noise that will still look like a trend. Publishing the score next to the trend is what stops a team spending a quarter chasing a movement that was never real.

From your own audience definition and from observed search demand, rather than from a brainstorm. You describe your ideal customer profile once, prompts are generated from those personas and from real query data, and they are suggested in bulk for you to edit, keep or reject. The reason is straightforward: a prompt list written from inside a company tends to ask the questions that company already wins, which produces a flattering baseline and no information.

You can, and you can paste your own list in. The problem is not the effort, it is the selection: a list written from inside a company asks the questions that company already answers well, so the baseline comes back flattering and tells you nothing you can act on. Generating from your audience definition and observed search demand puts the questions your buyers ask into the set, including the ones you lose. You still edit, keep or reject every suggestion, and the quality score will tell you which of your own prompts are not discriminating you from competitors.

Four axes on a 0 to 1 scale. Discriminance, weighted highest, is how far your mention rate diverges from your competitors' on that prompt: a question everyone wins tells you nothing. Stability is the variation in its daily mention rate. Actionability is whether the prompt is connected to a recommendation anyone actually followed. Recency is whether it has been probed in the last fortnight. Below three extracted probes the score is withheld rather than estimated.

Because they answer different questions. With search enabled, the engine retrieves current pages, so the result reflects your content and your crawlability. With search disabled, it answers from training data, so the result reflects what the model already believes about you. Strong with search and weak without means you are being retrieved but are not established in the model. The reverse means you are established but not being retrieved, which is usually a robots or extractability problem.