Does it tell me whether AI recommends a competitor or just names them?

What matching records, and what it leaves out

For each brand in the workspace, detection looks for its aliases at word boundaries in the answer text, and its owned domains among the citation hosts. Word boundaries mean "Acme" is found in "Acme is the safer pick" and not inside "Acmeville". What gets stored is a count, a position for the first mention, and which cited pages belong to whom. The sentence around the name is kept, because the whole answer is kept, but it is not scored.

So "we recommend Zenith", "Zenith is more expensive than Acme", and "avoid Zenith" each add one Zenith mention at the same position. The number does not know which of the three it is looking at.

Why we keep it that way

A sentiment score is a model's opinion about a sentence, and it drifts: the same answer can be read as positive on one day and neutral on the next, which would put movement in the series that never happened in the answers. A name is either in the text or it is not, and that stays true no matter how many times it is re-read. It is also what makes re-scanning safe: change an alias and every past answer is re-read with the same result it would have had on the day.

Where the judgement happens

Every answer is stored in full, with the model that wrote it, the date, its citations, and every brand mention highlighted. When a competitor's share moves on a prompt, our team opens those answers and reads what the engine is saying about them, which is the only reliable way to tell a recommendation from a warning. That reading is what our team brings to the monthly call, not a column in the table.