What AI says about your hotel: sentiment analysis
Being named by an assistant is only half the result. The other half is the sentence around your name, and in our own measurements the problem was almost never a bad description, it was a description with nothing in it.
AI sessions measured across the sites we track, last 30 days.
Quick answer
Sentiment analysis for AI answers records how an assistant frames your property when it names it, not merely whether it names it. In our own measurement pass of 3 and 4 September 2026, 57 questions across 5 AI surfaces produced 77 appearances across the measured travel and hospitality domains, of which 43 were framed positively, 34 neutrally and 0 negatively. The number to act on in that split is the 34.
Neutral is the result worth worrying about
Hotels brace for the negative mention and it is not what the measurements show. Across the 77 appearances of the September pass, none were negative and 34 were neutral, which is 44 percent of every time an assistant wrote the property into an answer.
A neutral mention means you were listed and nothing more. The traveller reads your name in a row of names with no reason attached to yours, which in a four line answer is close to not being there. It is also the most fixable of the three outcomes, because it is usually caused by pages that describe without ever committing to a comparable, checkable fact.
Sentiment also splits hard by assistant, which is why one global score would be misleading. Of the appearances measured in September, Gemini framed 12 of 15 positively, while Claude framed only 4 of 16 positively and 12 neutrally, on the same questions and the same day.
Sentiment is worthless if the mention count is wrong
Sentiment is computed on top of appearances, so it inherits every error underneath it. In that same September window a bug of our own returned 57 of 57 Gemini measurements as does not mention you. It failed silently, wrote nothing to a log and read perfectly well as a finding about a difficult surface.
After the fix, the same questions on the same day returned 15 visible and 15 cited, and 12 of those 15 were framed positively. Had we scored sentiment on the broken pass, the friendliest surface of the five would have been reported as the one that ignores you.
The rule that came out of it is written into the product. A cell that was not measured is painted differently from a cell where the property was not mentioned, with the reason attached, and a surface at a closed zero is treated as a measurement suspicion before it becomes a slide.
How it works in reporte.ai
Each measurement stores the answer it came from, so the framing is read against the actual sentence rather than inferred from a score. Runs are kept, which means you can compare this week against last on the same question list instead of restarting the count every time you open the page.
Sentiment is reported per surface and per period, never as a single global number, for the reason above: with 10 to 22 appearances per surface out of 57 questions, one average across five surfaces would describe none of them. One pass is not a verdict either, so a single reading proves nothing and the tone is compared across weeks. Where the sample is too short to conclude, the panel says so before it shows a figure.
Next to it sit the measurements it depends on: whether you are named and cited at all, who is described better than you and whether your pages are readable by an assistant in the first place.
FAQ
It is how the assistant frames you in the sentence that names you, not a score scraped from reviews. In our pass of 3 and 4 September 2026 the 57 questions produced 77 appearances across the measured domains, of which 43 were framed positively, 34 neutrally and none negatively.
It is the result with the most room in it. Of the 77 appearances measured in September, 34 were neutral, which means the assistant listed the property without giving a traveller a reason to choose it. A neutral mention is a page that states facts and never states a judgement worth quoting.
It can, and in that pass it was not: zero of the 77 appearances were negative. That is a useful baseline rather than a guarantee, because it tells you that what needs fixing today is the 34 neutral mentions and not a reputation fire.
Sharply. Of the appearances measured in September, Gemini framed 12 of 15 positively while Claude framed only 4 of 16 that way, with 12 neutral. Same property, same questions, same day, two very different descriptions.
By painting not measured differently from not mentioned, always. A bug of our own once returned 57 of 57 Gemini measurements as does not mention you, and after the fix the same questions returned 15 visible and 15 cited. Had we averaged sentiment over that, the surface would have looked hostile when it was silent.
More than a handful, and the panel says so before it shows a number. Per surface, the September pass produced between 10 and 22 appearances out of 57 questions, which is why sentiment is read per surface and per period rather than as one global score.
The evidence on the page the assistant is reading. Neutral framing usually comes from pages that describe amenities without a comparable fact, so the fix is a stated, checkable claim with a number, a date and a source, which is also what makes a sentence quotable in the first place.
No. Review sentiment summarises what past guests wrote. This measures what an assistant tells a future guest who has not booked yet, on questions where your name may not even come up. Both matter and only one of them is currently being measured by most hotels.