What we tested
We ran a visibility measurement for a brand that does not exist. No website, no customers, no coverage - a name invented for the test. The expected result was zero: a model cannot mention a company it has never encountered.
The measurement came back at 36% AI visibility.
Where the 36% came from
Breaking the answers down, every single mention came from prompts in which the brand name had already appeared. The model was repeating the name in order to say it knew nothing about it - sentences along the lines of 'I have no information about this company'.
Counted as a mention, that is a mention. Counted as visibility, it is the opposite of visibility.
Separated, the same measurement reads: 0% general visibility, 90% brand visibility.
Why this matters beyond one odd test
Prompt sets are usually built by someone inside the company, and people write what they know. Roughly half the prompts in a first draft tend to contain the brand name, because that is the shape of the question the author has in mind.
Every one of those prompts pushes the score up regardless of whether anything improved. The metric rises across reporting periods, nobody can point to what caused it, and the agency has no incentive to look closely.
What we do instead
We split the two from the first measurement and never merge them. General visibility answers 'do models bring you up when nobody mentioned you'. Brand visibility answers 'when somebody already knows your name, what do models say about you'.
Both are useful. They are also completely different questions, and only the first one tells you whether the channel is working.