Mention is not citation. Citation is not a recommendation.

A lot of AI visibility work collapses four different claims into one dashboard cell.

A mention only proves your name showed up.
A citation only proves a source was referenced.
A recommendation only proves preference language.
A conversion only proves a buyer acted.

Those steps can sit next to each other in a log and still not prove each other. Adjacency is not inheritance of proof.

I have been treating this as an evidence discipline problem, not a tooling problem.

What I do now on a short buyer prompt panel:

1. Freeze the exact wording of the question.
2. Note the engine and the date.
3. Log mention and citation as separate columns.
4. Only mark “recommended” if the answer uses preference language, not just a link.
5. Refuse to upgrade certainty when the next hop has no evidence of its own.

Ghost citations make this concrete. A model can reference a page without naming the brand. A brand can be named without being preferred. A preferred answer can still not convert.

If you only track one blended “AI visibility” score, you cannot tell which hop broke.

The useful weekly question is not “did we show up?” It is “what is this artifact entitled to prove?”

Curious how others are separating those hops in practice, especially after an answer fans out into sub questions the buyer never typed.

I put the logging workflow into The Money Prompt Lab if anyone wants a clipboard for buyer questions, opened or cited pages, and before/after checks. Happy to keep the method talk in the thread either way.