Method
Prompt-mean rule
The prompt-mean rule is Zebora's decision that every prompt contributes exactly one unit of weight to a score, however many times it was sampled. A question measured four times counts once, as the average of its four answers, rather than four times over. It is what keeps repeat sampling from distorting the result.
The problem it solves
Sampling is never perfectly even. A run fails, an assistant times out, a prompt is added mid-period. Without a rule like this, whichever questions happened to be sampled most would carry the most influence over the score — and whichever assistant completed the most runs would quietly override the weighting we configured for it.
What it means when you read a score
The figure answers a question about prompts, not about answers: on what share of the questions we track did the assistants mention you. That stays a stable thing to compare across periods even when the sampling underneath was uneven.
A prompt that ran three times, not four
Three questions are tracked across a month. One is mentioned in three of its four runs, one in none of its three runs after a failure, one in all four. Under this rule each contributes its own average and the tag score is the mean of the three — the missed run changes nothing but that prompt's own average. Count every answer equally instead, and the prompt that happened to run more would quietly weigh more.
Related terms
- Weekly visibility pool — The repeat sampling this rule exists to keep honest.
- Importance weight — The other thing deciding how much a prompt counts — importance, not frequency.
- Effective sample size — Why weighting a sample costs some of its statistical strength.
Last reviewed . Definitions are reviewed quarterly, and whenever the underlying measurement changes.
Don't let AI decide your brand's future without you.
See exactly where you stand vs. competitors—and what to do about it.