Intelligence
Guest sentiment analysis: finding the complaint hiding in your good reviews
A four-star review can contain a worse problem than a one-star one. Reading by score will never show you that.
Justin Ciappara
Co-founder, Revify8 min read
Here is a review a hotel might receive: “Beautiful room, lovely staff, great location. Breakfast was chaotic and we gave up queueing. Four stars.”
Filed by rating, that is a good review and it disappears into the pile. Filed by subject, it is the fourteenth complaint about breakfast this quarter, from a guest who otherwise loved the place. The second filing is the one that changes what you do on Monday.
The limit of a star rating
A star rating compresses an entire stay into one number. That is useful for the traveller choosing between options and nearly useless for the operator trying to improve. Two problems follow.
Averaging hides the specific. A property sitting comfortably at 4.3 is not 4.3 at everything. It is probably 4.7 on rooms, 4.6 on staff and 3.1 on one thing that never quite gets fixed, and the average is doing an excellent job of concealing which.
Severity is not frequency. The angriest review is usually about something that happened once. The expensive problem is the mild recurring irritation that a hundred guests mention in passing and never quite score you down for, until they book elsewhere next year.
What aspect-level sentiment actually means
Ordinary sentiment analysis gives a review one score: positive, negative, somewhere between. For the review above that produces something around neutral, which describes nothing that happened.
Aspect-level analysis does it per subject. The same sentence produces:
| Aspect | Sentiment | Evidence |
|---|---|---|
| Room | Positive | “Beautiful room” |
| Staff | Positive | “Lovely staff” |
| Location | Positive | “Great location” |
| Food and beverage | Negative | “Breakfast was chaotic, we gave up queueing” |
Do that across every review on every platform and you stop having a rating and start having a list, ranked by how often each thing is raised and which way it is moving. That list is the actual output. The score was only ever a symptom.
A worked example
A seaside hotel is steady at 4.2 for eighteen months. Management assumes the rooms need refurbishment, because that is the obvious candidate and the rooms are getting on. By aspect, the quarter looks like this:
| Aspect | Mentions | Negative share | Trend |
|---|---|---|---|
| Rooms | 210 | 12% | Flat |
| Staff | 180 | 6% | Improving |
| Breakfast | 96 | 61% | Worsening |
| Check-in | 74 | 38% | Worsening |
| Pool | 58 | 9% | Flat |
The rooms are fine. Guests mention them constantly and are largely content. The money was about to go to the wrong place. What is actually degrading is breakfast and check-in, and the two are probably the same problem, since both cluster between eight and ten in the morning and both smell like understaffing at a specific hour rather than anything structural.
That is a rota change and a second service point. It costs a fraction of a refurbishment, and it is invisible from the star rating, which has not moved because the rooms and staff are quietly carrying the score.
Turning themes into decisions
- Rank by mentions, not by anger. The loudest review is rarely the most expensive one.
- Look for time clustering. Complaints that bunch at a particular hour, day or season are staffing problems. Complaints spread evenly are structural.
- Check it against something you can see. If breakfast is the theme, stand in the room at half past eight before spending anything. Sentiment tells you where to look, not what you will find.
- Fix one thing at a time and watch that aspect. The value of per-aspect tracking is that you can tell whether the change worked, which a headline rating will never show you.
Where this goes wrong
- Treating the output as a verdict. It is a ranked list of candidates. Cost, feasibility, and whether the aspect actually influences bookings are still judgement calls.
- Too few mentions. Three negative comments about the lift is not a trend. Wait for the theme to repeat before you act on it.
- Ignoring the platform mix. Booking.com reviews come only from confirmed stays and skew differently from open platforms. A theme appearing on one source and nowhere else deserves suspicion.
- Reading it daily. Sentiment moves slowly. Checking it every morning produces reaction to noise; monthly produces decisions.
Why it matters more across several properties
For one hotel this is a useful habit. For a group it is the only practical way to compare, because a portfolio’s ratings are not comparable. A city hotel and a resort face different guests with different expectations, and a league table of star ratings mostly ranks property type.
Aspect-level data survives that. “Check-in is the worst aspect at four of our six properties” is a group-level finding you can actually act on, while “the Sliema property is on 4.1 and the Gozo one is on 4.5” tells you almost nothing about either. More on where this sits in a working process in what review software actually does.
Common questions
What is aspect-based sentiment analysis?
It scores opinion separately for each thing being discussed rather than for the review as a whole. One review saying 'lovely room, freezing pool' produces a positive score for rooms and a negative one for facilities, instead of a single neutral average that describes neither accurately.
Why is a star rating not enough to manage a hotel?
Because it averages away the detail that tells you what to do. Two properties on 4.2 can have completely different problems, and a property whose score is steady can be steadily hiding one recurring failure that costs it the guests who care about that thing.
How many reviews do you need for sentiment analysis to be useful?
Enough for a theme to repeat. In practice a few dozen mentions of the same aspect. Below that you are reading anecdotes. The useful signal is 'breakfast has been mentioned negatively fourteen times this quarter', not one strongly worded paragraph.
Can sentiment analysis tell you what to fix first?
It can rank candidates, not make the decision. It shows which aspects are mentioned most and trending worst, which is usually different from what management assumes. Cost, feasibility and how much each aspect actually influences booking decisions are still human judgements.