Sentiment analysis on AI answers is easy to sell and easy to misread. Green means good, red means bad, and a dashboard full of green feels like progress.
The trouble is that a model rarely says anything straightforwardly bad about a brand. It says something worse: something faintly qualified.
"Acme is a solid choice if you want convenience over range" is neutral by any classifier you point at it. It is also a sentence that will lose you customers who care about range, and it will keep losing them every time that question is asked.
Nothing in that sentence is negative. Every word of it is a positioning decision the model made on your behalf.
The useful signal is not the polarity, it is the vocabulary. Which qualities get attached to you when a model has to describe you in one clause?
If the same three adjectives come back across dozens of different questions, that is your position as the model understands it. It may not be the position you have been paying to establish. Finding out that the word attached to you is "affordable" when you have spent two years building a premium story is the sort of thing worth knowing before the next campaign, not after.
A single brand's sentiment drifting week to week is mostly noise — rephrase the prompt and it moves. What is stable, and therefore worth tracking, is the gap between how you are described and how the brand next to you is described in the same answer.
If you are consistently the sensible option and they are consistently the exciting one, that is a real finding, and it does not depend on any of the underlying scores being precise.
Sentiment tells you where to look, not what to do. A qualification that keeps recurring is a claim the model has found support for somewhere. Usually that somewhere is a forum thread, a review page, or your own copy read less charitably than you intended.
Chase the source of the adjective. Changing the adjective means changing what there is to read, and that is slow work — but it is at least work you can actually do, which is more than can be said for trying to argue a score upwards.
Contents
From tracked prompts to a score you can act on — the mechanics.
Read →Not your homepage. A working theory of the sources that shape what AI answers say about a brand, and what that means for where you spend effort.
Read →What the lines mean, which movements are real, and the three ways this chart gets misread.
Read →Tell us your domain and we'll send you a visibility report — free, no card.