What an honest AI visibility baseline actually contains

Most AI visibility reports we are shown by prospects fail the same test: there is no way to run them again. Without that, nothing measured afterwards can be compared to anything, and every claim of improvement is a story.

Five things make a baseline re-runnable.

Written down, versioned, and not edited between runs. The moment a prompt is reworded because it produced a disappointing answer, the panel stops measuring the market and starts measuring the person writing the prompts.

One blended visibility score can stay flat while you lose the engine your buyers use and gain the one they do not. Report the five separately or do not report 1.

Prompts where nobody was named, and prompts where you were not, both get written down. A report containing only the favourable runs is not a baseline, and it is the single most common thing we find in the decks prospects bring us.

Does a link in a source tray count as a mention? Does being named third count the same as being named first? There is no universal answer. There is only a rule that is written down before the run and applied to everyone in the category equally.

A baseline with no scheduled second run is a snapshot. Answer engines change their retrieval behaviour without announcing it, so a number with no date on it stops being true quietly.

None of this is difficult. It is just auditable, which is a different thing from impressive, and it is the reason our tier pages publish the panel size instead of a visibility score.