“Northfield Heating & Air is known for clean, careful work, but wait times can run up to eight weeks.” One sentence from an AI answer, with a recommendation and a brake. Caveats like this aren't random. They come from sources the model has read, and they keep coming back as long as those sources say so.
Where sentiment comes from
AI models don't form their own opinions. They summarize what they read: reviews, forum threads, comparisons, news articles. If three of the five sources they read mention the wait time, the wait time ends up in the answer. That makes sentiment something you can influence—but only if you know which source the caveat comes from.
How to measure it
For each mention, a language model uses a fixed rubric to rate the tone as positive, neutral or negative, and records the source sentence as evidence. Across all mentions, that produces a score (0–100) and a distribution. More important than the score are the recurring themes: price, quality, service, delivery. A theme that shows up in six negative mentions a month is a concrete problem with a concrete cause.
The source sentence as evidence
A sentiment score without the source sentence is an opinion about an opinion. That's why we show the sentence where it happened, the AI model and the date with every rating. This has two benefits. You can correct a wrong classification. And in a team meeting you can show what AI literally says, instead of arguing about a number.
What to do about it
- Find the source of the caveat (often a review platform or forum) and respond there with substance.
- Fix the underlying problem if it's real—otherwise AI will keep repeating it.
- Publish current, concrete information on your own site (realistic lead times, for example). Models like to cite the first party.
- Measure the effect after thirty days: does the theme disappear from your mentions?
Sentiment versus competitors
Sentiment is relative. A score of 71 is good if your competitors sit at 58 and 62, and mediocre if they're at 80. So don't just compare your own trend. Look at which caveats AI attaches to your competitors too. Those are exactly the points where you can stand out in your content and your sources.
AI doesn't give stars. It gives a clause. Measure that clause, find the source, fix it and measure again.