Replying to hotel reviews with AI: what it saves and what it costs

The tools that write your review replies for you work, and they work well. You paste the review, out comes a courteous, correct paragraph in the right language. Forty outstanding reviews get cleared in an afternoon.

The question almost nobody asks is the second one: what does that afternoon cost you?

I spent a day on the research that has tried to measure it. The honest summary is that one thing is solid, one is contested, and a third has not been measured by anyone yet. Here they are in that order, each with the place where it is weak.

The solid part: replying pays

The best evidence out there is from 2017, published in Marketing Science by Davide Proserpio and Georgios Zervas. They compared Texas hotels across two platforms: Tripadvisor, where managers reply often, and Expedia, where almost nobody does. The comparison separates the merit of the reply from the merit of the hotel: if a property genuinely improves, it improves on both; if it improves only where it started replying, the reply is what did it.

The result: a 0.12-star increase in rating and 12% more reviews for hotels that start responding.

There is a side effect the authors state plainly. Hotels that start replying get negative reviews that are fewer but longer. Their explanation is that an unhappy guest, knowing someone reads and answers, thinks twice about the throwaway complaint they could not defend. The ones who still write, argue their case.

0.12 stars is not as small as it sounds

On a one-to-five scale, 0.12 looks like nothing. But hotel ratings are not spread across the scale: they are bunched at the top, in a narrow band. And platforms round to the nearest half star.

That is where the decimal becomes visible. A hotel at 4.24 that reaches 4.25 does not gain a hundredth: it gains half a star in the little circle a guest sees in a list, before opening the page. In that study's data, 27% of hotels that started replying gained at least half a rounded star within six months of their first response.

So far, everything argues for replying to everything, fast. Which is exactly the point where the automatic tool looks like the right answer.

The contested part: replies that resemble each other

In 2021, in Tourism Management, Shan Liu, Na Wang, Baojun Gao and Michael Gallivan published a study on something narrower: not whether a manager replies, but how much their replies resemble one another. They call them rote responses, and they measure the similarity by comparing a single hotel's own texts. The material is 334,671 reviews and 169,794 manager responses from 868 hotels on Tripadvisor.

What they observe: where a manager's replies resemble each other more, the following month brings fewer positive reviews and a lower average score. The drop in volume is in the enthusiastic reviews, not the negative ones: template replies do not bring more criticism, they bring fewer compliments.

And one detail changes who the warning is addressed to: the effect is stronger for independent hotels than for chains. The authors close by recommending that replies be differentiated "particularly for independent hotels".

There is a logic to it. From a chain, a guest expects a corporate voice, and a standard reply betrays no promise. From a family-run property they expect a person: that is why they chose you over the chain. When they find the corporate voice, what they discover is not that you use a tool. It is that the difference you were selling was not there.

But this is an association, not an experiment. The data is observational: nobody drew lots to decide which hotels would reply by template. The confound is obvious, and a one-month lag does not solve it: a hotel that is letting things slide does both at once, copies and pastes its replies and lets the service slip. You can say "goes together with", not "causes".

Not every study says the same thing

This is the part that usually disappears from articles on the subject, and without it none of the rest makes sense.

Across 332 Chicago hotels, Gong and colleagues (International Journal of Hospitality Management, 2022) found that replying with official, structured text raised both review volume and rating. The opposite side of the argument from the study above.

On Booking.com, in Belgium, Lopes and colleagues (Journal of Service Management, 2024) looked at something none of the others look at: actual bookings, not star ratings. Across 18,320 reviews and 5,564 replies from seven hotels, personalising the reply did not move future bookings. Apologising reduced them by 4.9%.

And in the Journal of Marketing Research (2018), Wang and Chaudhry found that tailored replies amplify the benefit when you answer a negative review, but also amplify the damage when you answer a positive one.

Taken together, these three do not demolish the first one. They say that "always personalise everything" is not a conclusion the research supports. They say what matters is which review you are answering.

Why an Italian hotel should read all this carefully

Three limits, and they are big ones.

Almost all of it comes from Texas, on Tripadvisor. The 0.12-star study uses Texas hotels. The rote response study, 868 hotels on Tripadvisor. A small Italian property lives on Booking and on Google, where only people who actually stayed can review, and where the population of reviewers is a different one. The only Booking study among those I have cited is also the only one that finds no effect from personalisation.

Those hotels have far more reviews than yours. In the rote response sample each property averages a few hundred reviews. If you collect twenty or thirty a year, "next month's volume" is not a signal for you: it is noise.

The data stops in 2017, and I read the abstract, not the article. It sits behind the publisher's paywall, there is no open access copy, and I have not seen the tables. The numbers and directions above come from the abstract and the public fragments. I would rather write that than pretend I read the paper.

The part nobody has measured: AI

One thing needs saying plainly here. Neither of the two main studies is about artificial intelligence. They measure replies that resemble each other, not generated replies.

And the inference "AI equals template replies" can be turned on its head, with good arguments. The similarity being measured is between one hotel's own replies: a well-prompted model produces texts less similar to one another than a human copy-paste, not more. What that study penalises is repetition, and repetition is a choice made by whoever replies, not a property of the machine.

What has actually been measured about AI is subtler, and concerns two different things.

In Cornell Hospitality Quarterly (2024), Litvin and Tan asked participants to tell a manager's own reply from a generated one: they could not do it reliably. In another study by the same authors in Tourism Management, however, when participants knew the reply was generated, they liked it less.

The combination is worth more than a thousand opinions: the problem is not that the guest notices, because they do not. The problem starts if you tell them, or if they find out.

What I would do

Not "don't use AI". That would be a comfortable, useless piece of advice, and the time it saves is real, as it is for the chatbot that answers questions before the booking.

Three things instead, and they hold even if you throw out the most contested studies.

Use it for the draft, not for sending. The documented risk is not having used a tool, it is the results resembling each other. If you read and change, the resemblance breaks: thirty seconds per reply.

Put one thing in every reply that is true only of that review. The room name, the dish they mentioned, the storm that weekend. It is the detail no tool can invent without knowing it, and it is the difference between a reply and a receipt.

Change the opening above all. Ten replies that all start with "Dear guest, thank you for your valuable feedback" are not ten replies: they are one reply printed ten times, and whoever reads your profile sees them stacked in a column, one under the other.

And on Tripadvisor there is also a rule

One last thing, which is almost always reported backwards. It is worth reading at the source, because it takes two minutes.

In its management response guidelines, Tripadvisor writes: "we ask that you abide by Tripadvisor's Content Policy and the following rules for management responses to reviews". It asks you to abide by the Content Policy, and the link goes to the general content guidelines page.

There it says: "Review content generated by artificial intelligence (AI) is not permitted and a violation of Tripadvisor guidelines. Users found to be submitting AI-generated text and/or images to the site may be banned, in addition to having their content removed".

The first sentence is about reviews. The second is not: it is about users submitting AI-generated text or images to the site, and it says they may be banned on top of having the content removed.

In fairness, the other side needs saying too: Tripadvisor has never published a rule specifically about AI in management responses, and when it publishes removal figures it talks about traveller reviews. In its transparency report of 18 March 2025 it states it removed "over 214,000" reviews it had reason to believe contained AI-generated text, across 101,411 properties in 189 countries, against 31.1 million reviews received. That is cautious wording and it deserves to be reported as such: these are not reviews confirmed to be artificial, they are reviews suspected of containing artificial text.

Between "it is not forbidden" and "there is no dedicated rule, but the general one applies to you" there is a difference worth knowing before you hand a tool every reply your property will ever send. Checked on 14 September 2026: pages change.

If you are working out where to put AI in running your property, the picture of what European hoteliers actually use helps you decide where to start. Reviews are one of the places where the time saved is most real, and also the one where the shortcut shows most.