Success Story: AI Review Analysis Across Booking.com, Google and Tripadvisor for a Group of Five Hotels

At a glance

• Five hotels, four platforms, one panel: Booking.com, Google, Tripadvisor and Hotels.com, collected automatically every month
• Reviews in any language, conclusions in the manager’s language (Spanish, English, French or German)
• Every conclusion linked to the reviews behind it, and checked by a second AI model before anyone sees it
• A 60-word monthly summary with memory: what has been fixed, what keeps coming up and what is new compared with previous months
• Competitors analysed the same way, to compare scores and what their guests praise and complain about
• Around 20 hours a month of manual work saved (estimate): half a day per hotel, plus half a day for each competitor monitored

The client

A hotel group with five hotels. Each one is listed on Booking.com, Google, Tripadvisor and Hotels.com, and between them they receive hundreds of reviews a month, in many languages.

Management wanted to know, every month and for each hotel, what guests praise and what keeps going wrong, without anyone having to read every review.

The challenge

• Twenty places to look. Four platforms for each of five hotels, each with its own panel, its own score and its own way of showing reviews.
• Many languages. German, English, French, Dutch, Italian, Spanish… Nobody reads them all, so the reviews in other languages simply went unread.
• A score says how you’re doing, not why. An 8.4 on Booking.com doesn’t tell you whether the problem is the breakfast, the noise or the air conditioning.
• Anecdotes instead of the big picture. Reception answered reviews one by one, but spread across many shifts, nobody saw them all together. Patterns got lost, and whoever read the latest reviews remembered the loudest complaint, not the most frequent one.
• No memory. Was the breakfast complaint from June solved? Nobody could say without rereading June’s reviews.

What we built

Automatic review collection

Each hotel is set up once with its listing links. From then on, the platform collects its reviews from the four platforms every month, through review data APIs and our own scrapers, removes duplicates and stores them with their score, date, platform and text, including the separate “liked” and “disliked” fields that Booking.com uses. Previous months can be loaded too, so the history is there from day one.

Collection runs as a job that saves its progress at each step. If something fails halfway, for example a platform not responding, it resumes from where it stopped instead of starting over.

One panel per hotel, and the group at a glance

• The group’s hotels in a sidebar, each with its current score
• Competitor hotels, added and analysed just like the group’s own: their score month by month, and what their guests praise and complain about
• Score by month, overall and for each platform, to see the trend and not just this month’s number
• How many reviews came in, per day and per platform
• Every review in a table, translated into the manager’s language when it’s written in another one
• Export to Excel of the month’s reviews and conclusions

AI conclusions you can verify

Every month, AI reads all of a hotel’s reviews, in whatever language, and extracts the specific feedback: what guests liked and what they didn’t, grouped by topic (reception, breakfast, noise, cleanliness, parking…). If several guests complain about the breakfast, one about the timing, another about quantity and another about variety, that becomes one conclusion that mentions all three, not three separate lines.

Each conclusion comes with:

• An importance score from 1 to 10, based on how often it comes up, how serious it is and how actionable it is: a small balcony is hard to change; a bathroom that wasn’t clean is not.
• The reviews that support it. A click shows exactly which guests said it, on which platform and when.

Before a conclusion reaches the panel, a second AI model from a different provider checks it against each review it cites. Reviews that don’t really support it are removed, and a conclusion left with no supporting reviews isn’t shown. The manager can trust that what they read is in the reviews and isn’t something the AI made up.

A monthly summary that remembers the months before

On top of the conclusions, AI writes a summary of the month in 60 words at most, in the manager’s language: one or two things guests loved, and above all what needs attention, in order of priority.

To write it, it receives this month’s conclusions and the summary of the previous month (if a month is missing, it looks back up to four months). That way it can say:

• what has been resolved (“complaints about bathroom cleanliness have practically disappeared”),
• what persists (“the terrace noise is still the main issue”),
• what is new (“a new problem appears with the air conditioning in the upper rooms”),

and it avoids repeating last month’s wording. Because each summary builds on the one before, problems that drag on are flagged as such, and improvements are noticed, without anyone comparing months by hand.

When the month’s analysis is ready, the manager receives the summary by email, with a link to the full panel.

How we worked

We started with the group’s real reviews. We compared several AI models on the same months of real reviews before choosing one for analysis and another for checking, and adjusted the instructions with management until the conclusions said what they needed to hear: grouping similar complaints instead of listing each variation, and ranking by how actionable something is, not just by how often it appears.

We also log the cost of every AI call and every review collected, so we know exactly what each hotel’s monthly analysis costs.

The results

• One place for five hotels and four platforms, instead of twenty panels.
• From anecdotes to the full picture. Management now knows where each hotel’s biggest problems are, and which strengths to keep looking after and building on.
• Management reads conclusions, not reviews. The month’s summary takes a minute to read, and the details are one click away.
• Every review counts, in every language. The German and Dutch reviews weigh as much as the Spanish ones.
• Conclusions they can trust. Each one shows its reviews, and a second model has checked them.
• Month-to-month follow-up. They can see whether what they changed worked, and what keeps coming back.
• A clear view of the competition. They can see where nearby hotels do better or worse, and why, according to their own guests.

Estimated time saved

Doing this by hand properly, reading every review of the month on four platforms, noting what each one is about, drawing conclusions and writing up a summary, takes about half a working day per hotel every month. For five hotels, that’s around 20 hours a month, and another half day for each competitor they want to study, so the figure grows with every hotel they monitor. With the platform, management spends about an hour reading the summaries and checking the details.

In practice, nobody had those hours, so the work simply wasn’t done: what reached management was anecdotal. Now it gets the whole picture every month.

Key takeaways

1. Bring everything together before analysing it. Half the value is simply having every review from every platform in one place.
2. Make the AI verifiable. Linking each conclusion to its reviews, and having a second model check them, is what makes a manager act on it.
3. Give the AI memory. A summary that knows what happened last month is a follow-up, not just another report.

Want to know what your guests are really saying?

We can build this kind of analysis for your hotels, or adapt it to your own reviews and surveys. See also our reputation management service and AI for hotels. Book a free discovery call.