In AI visibility, sentiment is how favorably an answer frames a brand when it mentions it: recommended, neutral, caveated, or warned against.
A mention is not automatically a win. "X is the market leader, though users report billing issues" and "X is the best choice for most teams" are both mentions of X with very different value. Sentiment analysis classifies that framing at scale.
Sentiment usually traces to sources. Models echo the tone of what they read. Persistent caveats in AI answers tend to have a findable origin in reviews, forums, or comparison articles.
Why sentiment is not reported as a number here
Framing is real and it matters, but collapsing it to a score invites false precision. "Best for most teams", "the market leader, though pricing is opaque", and "widely used" are three different endorsements, and any single number that separates them is a judgement dressed as a measurement.
The honest unit is the sentence. Reading the recorded answer tells you how a model framed a brand more reliably than a sentiment score derived from that same sentence would.
Tracing a caveat to its source
Persistent caveats usually have a findable origin. If several models hedge on the same dimension — support, pricing transparency, migration difficulty — that hedge is almost always echoing something specific and locatable in reviews, forums, or comparison articles.
This makes sentiment work concrete rather than atmospheric: find the source, and you know what would have to change for the framing to change.