Named rate

A prompt is run many times across several engines. Each returned answer is checked for the brand name using deterministic word-boundary matching, so the result never depends on a model deciding whether a mention counts. The named rate is the number of answers that matched, divided by the number of runs that actually completed.

A run that failed, for example an engine timing out or returning an empty response, is excluded from the denominator entirely. It is never counted as an answer that ignored you. Counting failures as absences would depress every brand's rate and make any tool reporting it look more useful than it is.

This is the distinction most visibility scores erase. An answer can name your brand while recommending a competitor, name you as the expensive option, or name you in a list of things to avoid. All three count as named.

That is why endorsement is graded in a separate layer rather than folded into this number, and why that grade is only recorded when the classifier can produce a verbatim sentence from the answer that names you. If it cannot, the run stays unevaluated and shrinks the sample instead of quietly counting as not recommended.

A rate from a finite sample is an estimate, not a fact. Twenty runs that name you eight times give 40 percent, but the true rate could reasonably sit anywhere in a band around that. Every rate carries a Wilson confidence interval so the width of that band stays visible, and a change is not called a win while the before and after ranges overlap.