AI for hotels: can you trust it to write the facts about your hotel?

In our first test, AI got a hotel’s facts wrong in 38 of the 77 answers we checked: one in two. We wanted to know one precise thing. If a machine reads everything the web says about a resort and fills in its fact sheet, with transfer minutes, kids’ club ages, villa sizes and the seaplane fare, how much of that sheet is true?

On 3 October 2026 we held 7,050 answers about 33 Maldives resorts. A language model, the kind of program behind ChatGPT, had written them after reading official sites, booking portals, guides and reviews. We drew 80 of them, from 31 resorts, and gave them to four reviewers that did not know how they had been produced: four separate AI sessions, each starting from scratch and told to find mistakes. For every answer they downloaded the cited page again that day and opened the resort’s official site.

Three answers could not be verified. Of the other 77, 38 were wrong, and 22 were wrong in a way that would have changed a traveller’s choice: a 2023 fare given as current, an arrival cut-off that applies only to the seaplane, a residence recorded at half its real floor area.

The sample was built to test the method, so it deliberately held the facts that decide a booking and the ones on which sources disagree. The group closest to an ordinary fact sheet is the random one: 6 wrong out of 19. The measurement covers one destination, one day and our own tools, and it shows how often an automatic reading goes wrong when it has no rules. How often ChatGPT goes wrong when you ask it for copy about your hotel still has to be measured, hotel by hotel.

The job is the same, though. Anyone who asks an assistant for their hotel’s family page or transfer details is having it do, on a small scale, what we did: write facts from what it remembers and what it finds online. That text then goes on the site under the hotel’s name.

It happens outside travel too. A study coordinated by the EBU and led by the BBC, published on 22 October 2025, had journalists from 22 public broadcasters assess more than 3,000 answers from four assistants to questions about the news. In 45% they found at least one significant issue, and in 20% major accuracy problems, including invented details and outdated information.

AI makes up a fact because it writes by completing a sentence with the most likely words, and a likely sentence can be false. Ask it from what age the kids’ club is open and, if it does not have your club’s age in front of it, it writes a plausible one: the age at plenty of other kids’ clubs.

The companies that build the models say so themselves. Google’s guidance on generative AI content reminds publishers that these models predict words, and asks for a manual fact-check of every text before it goes out. OpenAI, in an article of 5 September 2025, explains why a model does not stop when it does not know: it is graded on tests where a guess scores better than “I don’t know”.

The same article says which facts are most at risk: rare, arbitrary ones that follow no rule. Its example is a birthday. For a hotel they are the minimum age at the kids’ club, the minutes by boat from the airport, the price of a babysitter.

Our check has a textbook case. For one resort’s teen club the sheet said “from age 16” and mentioned a PlayStation 4. The source gives ages 13 to 17, and the console appears on no page: the model had described a teen club the way teen clubs are usually described.

When it does have sources in front of it, AI takes wrong facts from old documents, from copies on portals and from pages with no date. In our check the most frequent error, 18 cases out of 38, was choosing the wrong source among several that said different things.

Here is one you can verify. The FAQ of Gili Lankanfushi, read on 8 October 2026, puts the resort 30 minutes by speedboat from the airport. A press release of January 2018, still online on the same site, says 20. They are two public pages of the same hotel, eight years apart. On 3 October the old number was repeated by six portal pages and by a travel agent’s PDF: seven voices against one. Our sheet could not decide and recorded the two numbers as a disagreement.

Of the 20 disagreements in the sample, only 3 were real. In 10 it was enough to take the official page or the most recent one. In the other 7 there was no disagreement at all: the same figure read two ways, two different villa types, one source in two languages.

Assistants pass such errors on: of 98 we found on the web for 11 hotels on 27 September 2026, they repeated 31. The count is in our guide on how to get your hotel on ChatGPT and Gemini. In our August 2026 study of answers from ChatGPT, Claude and Gemini, when an answer named one of the world’s 50 best hotels, the hotel’s own site was among the links shown 24% of the time.

A real quotation does not make the answer true. In our check, 144 of the 166 sentences quoted in support of the answers were on the downloaded page word for word. The errors came afterwards, on the way from the sentence to the fact.

Three cases from the check. A guide wrote that the dining gazebo of a residence faces the sunset: on the sheet, the villa faced the sunset. A resort’s site mentioned an “outdoor bathroom”: on the sheet it had become an “outdoor bathtub”. Another page said “partial ocean view”: the sheet said “ocean view”, the same as for the villas with a full view of the sea.

On 4 October 2026 we found a piece of evidence that proved the opposite of its claim. The line “guests praise the food” was backed by a review in Thai in which the guest wrote that the food had not suited them. When an assistant shows you its source, read the quoted sentence and check that it holds that very number or that very judgement.

AI most often gets wrong the facts that come with a condition attached: a price with taxes, a time that applies to one transfer mode, an offer tied to one rate. These are cases from our check of 3 October 2026, one for each kind of fact.

Dates are a case of their own. On 4 October 2026 a reviewer rechecked 20 years drawn at random: the opening year was right for 11 resorts in 12, and the year of the last renovation was wrong or uncertain for 4 in 8. In the one opening-year error, a reopening after a rebuild had been read as the resort’s first opening.

The errors came down when the reading started from the official site and was allowed to answer “they don’t say”. Our first remedy was a different one and it was not enough: questions split by subject, and stricter rules for choosing between sources, the most official first and then the most recent. That same day, 3 October, four reviewers checked 68 answers about 9 resorts read that way, and 48 contained an error. The sample was different and the brief stricter, so the two results cannot be compared like for like. It was enough to stop.

We then set the two readings side by side on three resorts. For each one, other AI sessions had prepared reference answers separately, reading the sources without seeing the machine’s work. Of 198 reference answers, the new reading got 160 right and the previous one 111.

The judge also found two errors in the references and three faults in our own checks. “1,800” was read as two numbers, 1 and 800. In “m2” the check saw the number 2. And a resort’s notice of works, repeated at the top of every page, was thrown out with the menus. That notice was still on the site of COMO Cocoa Island on 8 October 2026: arrivals and departures by the South Jetty until 15 October, while the North Jetty is being worked on. With the faults fixed, a second judge ran the comparison again.

The result is 1 serious error in 171 answers given: the 170 in the chart, plus one the references did not have. It was a “by speedboat” that had become “by speedboat from the airport”, when that resort is reached by a domestic flight and 25 minutes by boat.

The cost is in the 30 missing answers. The machine stayed silent even where an answer existed, 23 times on an official page. A cautious fact sheet has gaps, and somebody fills them by hand. Reading one resort’s official site in full cost us between $1.05 and $3 of computing, about $2 on average, with the site read 9 times: once for the list of villas and eight times by subject. The real expense is the checking.

To use AI without publishing mistakes, you give it the facts yourself, dated, and ask it to write with those alone. The method that worked for us can be copied with any assistant, ChatGPT, Claude or Gemini, in seven steps.

These are the same things the model builders recommend. Anthropic’s documentation on reducing hallucinations lists three techniques first: let the model say it does not know, have it extract word-for-word quotations before it writes, and have it withdraw any claim left without a quotation.

We applied this quoted-sentence rule on the night of 6 October 2026 to family facts at Maldives resorts: kids’ club ages, babysitting, teen programmes. A fact went in only if the quoted sentence was on the downloaded page word for word. That let in 201 facts, 145 from official sites and 53 from guides, and the kids’ clubs with their ages on record rose from 57 to 109 of 139.

In a separate article we measured how much the world’s best hotels publish, and whether a blog brings traffic. And our guide to who actually does the marketing in a hotel will help you decide who does the reading.

Never hand over to AI a fact the guest will use to decide or to pay: prices, taxes, ages, rules, dates. Drafting, summarising and translating a text that has already been verified can be handed over.

It is the hotel that answers to the guest for the mistake. On 14 February 2024 a Canadian small-claims tribunal ordered Air Canada to pay C$650.88 in damages to a passenger who had bought two flights trusting a wrong answer from the chatbot on its website. The decision in Moffatt v. Air Canada says a company is responsible for all the information on its website, whether it sits on a page or comes from a chatbot. It is the ruling of a tribunal in British Columbia and it applies there. Any guest understands the principle: what is on the hotel’s site, the hotel has said.

The empty field is the strongest temptation, for us as well. On 26 September 2026, rereading our own product’s instructions, we found one that asked the model to complete missing fields by searching the web, at a step where the model could not search: it would have completed them from memory. It had never fired, because the texts produced up to that day were complete. A gap is information: it says a source is missing.

The same rule holds for photographs. On 4 October 2026 we decided not to retouch real hotel photos with generative models, because they repaint the scene and add things that are not there. AI does judge a text well when the versions really differ: on 25 September 2026 a small model put four versions of 30 real adverts, the original and three copies made worse on purpose, in the right order 30 times out of 30.

Google’s guidance does not forbid text written with AI: it asks for it to be accurate and useful. Its page on generative AI content, updated on 1 October 2026, says generative AI is useful for researching a topic and giving structure to original content, and warns that producing many pages with no value for readers may break its spam policy on scaled content abuse.

The same page asks for the manual check on the lines nobody rereads as well: the page title, the description written for search engines, the structured data (the labels in the code that tell machines what a page contains), the alternative text of images. For a hotel that is where an invented “free parking” travels furthest, because those lines appear in search results and assistants read them too.

This is what we do ourselves: our hotel guide and our app for ChatGPT and Claude show next to each fact its source and the day we read it. Some errors remain, and with the date a reader can find them.

The checking was done by independent AI sessions, on public pages downloaded again that day: 149 cited pages and about forty more official ones. No person reread the 80 answers one by one. The coordinating session went back to the pages for the four verdicts that weighed most, and all four held; it corrected one other verdict. For this article, on 8 October 2026, we reopened the public pages we name.

The sample was not random: 30 facts that decide a booking, 20 disagreements between sources, 20 facts drawn at random and 10 answers marked “not found”, with no more than 2 per resort in each group. We counted an answer as wrong when its value was incorrect or out of date, when its quotation did not support it, or when it was given as unfindable although it existed.

The final comparison covers three resorts, a single run and a judge that was itself an AI session. That 1 in 171 is the result of that test. The error rate of our guide today is a different measurement, one we have not made. Everything here concerns Maldives resorts, read between 3 and 6 October 2026: a city hotel may give different numbers.