Rixe runs a defined set of category prompts through each engine's live interface once a day and parses the answer that comes back. For every run it records whether the brand was named, its rank among the brands mentioned, the sentiment of the sentence around the name, and every source the answer cited. No score is published from a single run: below five runs a prompt is marked insufficient sample, and every score ships with a Wilson confidence interval.
A single run proves very little, which is the part most measurement approaches skip. Generative models are probabilistic: ask the same question twice and the set of brands can change, the order almost certainly will. Publishing a score from one answer produces a number that moves for reasons that have nothing to do with your brand.
Rixe treats each prompt as a sample rather than a reading. Below five runs a prompt is marked insufficient sample and shows no score at all, and every score that does publish carries a Wilson confidence interval, so a 41% built from twelve observations is visibly less certain than a 41% built from ninety. When an engine is unavailable the data is flagged as missing rather than quietly counted as a zero — a broken adapter must never look like a drop in visibility.
The raw answer is stored verbatim alongside the parsed metrics. That is what makes a result auditable: when a number moves, you can open the answers behind it and read the sentences that changed, rather than trusting a chart.
The prompt set is the instrument
Everything downstream depends on the questions, so they are chosen before anything is measured and then left alone. They are written the way a buyer would type them — a sentence with a budget, a team size or a constraint in it — rather than as keywords, because that is what the engine is given in real use.
A set that is heavy on awareness questions and thin on comparison questions will flatter you, because broad questions produce long lists. Staging the set by where a buyer is — learning, comparing, deciding — is what keeps a healthy-looking score from sitting next to a pipeline that disagrees with it.
Repetition, and why one run proves nothing
Generative models are probabilistic. Ask the same question twice within an hour and the brands, the order and sometimes the facts can differ, which makes a single screenshot the least reliable artefact in this field. The answer is boring and it is the whole method: run the same prompt repeatedly, on a schedule, and report the rate.
Repetition is also what makes a confidence interval possible, and the interval is what stops a team acting on noise. Below a floor of runs per engine no score is published at all; the screen says the sample is too small instead of drawing a line that would be read as a trend.
What is recorded from every answer
Each answer is kept whole, then read for five things: whether your brand was named, where in the order it appeared, what was said about it, which pages the engine cited, and which competitors stood in the same sentence. Keeping the raw answer is the part that makes the rest auditable months later.
Tone is classified automatically, which means it is sometimes wrong, and a measurement platform should say so where you can see it rather than in a footnote. Anything that looks misread can be flagged, and the flag carries the evidence with it so the correction is about a specific answer on a specific date.
