Translating a hotel website with AI: which pages, who checks and how to publish

AI website translation means producing each language version of a hotel site from one source text with a model, then having a fluent person check the pages where a mistake costs money. Translate policies, prices, allergen and accessibility notes first, keep one source text, publish each language on its own address, and have counsel confirm legal pages.

A guest reads your cancellation policy in Russian, understands something slightly different from what you wrote, and argues about it at the desk. The translation was fluent. It was also wrong. Fluent and wrong is the most dangerous combination on a hotel website, because the sentence that reads best is the one the guest trusts most. This guide gives you a working method: which languages to pick, which pages to translate first, where models fail on hotel text, who checks what, and how to publish the versions so guests and search engines both find them.

Which languages should a hotel website be translated into?#

Start from your own reservation records, then check them against national arrival data. The languages your guests actually book in come first, not the languages a template happens to offer.

Pull the last twelve months of reservations and count guest nationality, or the language each booking was made in if your system records it. That list, not a market report, describes your hotel. A property in a ski resort and a city hotel in the same province can have completely different top three languages.

National data is a starting point only. The Ministry of Culture and Tourism's 2025 yearly border statistics bulletin (opens in a new tab) counts 52,775,055 foreign visitors arriving in Türkiye in 2025, with the Russian Federation at 6,903,981 (13.08%), Germany at 6,745,190 (12.78%) and the United Kingdom at 4,270,036 (8.09%). Say what that number is: nationality comes from passport police records at border gates, and the foreign visitor count includes same-day visitors, so it describes arrivals to Türkiye, not guests in your rooms. For more on reading this data, see our article on Türkiye tourism source markets 2025.

Source: Ministry of Culture and Tourism, 2025 yearly border statistics bulletin. National arrivals, not hotel guests. Public data. No endorsement implied.

Group where one language covers several markets, and keep English as the fallback for everyone else. A reader who finds no version in their language will use English if it is complete, so a complete English site beats four half-finished ones.

Check the operation, not only the website. A language you cannot answer the phone or the reservation email in creates a promise the front desk cannot keep. If nobody on shift reads Russian, a Russian website needs a plan for what happens when a Russian speaker calls. Our piece on a language plan for your guest mix covers that side.

Translate the top two or three languages first, then look at direct bookings and enquiries per language version before adding the next one. Do not add a language for the flag row. A stale version in a language nobody reads is worse than no version, because it looks maintained.

Which pages should be translated first, and which need a fluent check?#

Translate in order of what a mistake costs. The pages that decide money, safety and legal position come first and get a human check; descriptive pages can wait.

The first tier is money and safety: cancellation and deposit terms, what the price includes, child and pet policy, allergen notes on restaurant and room service menus, accessibility notes, check-in times and transfer instructions. The second tier is the sale: room types, board types, photo captions, spa and facility opening hours. The third tier is soft content: about us, destination guides, blog posts. These can go live from a checked model draft.

Legal pages, such as pre-contract information, the personal data notice and distance sales terms, are translated and then confirmed by counsel. The translation is not the legal text, and a model is not a lawyer. Confirm dates, penalties and your own obligations with a legal adviser, as with any compliance question.

A page that is wrong or vague in the source language will be wrong in five languages. Fix the source first, then translate. Give every first-tier page a named checker and a review date, written down where the marketing team can see it. Keep the page structure identical across languages so a reader who switches language stays on the same subject. Our hotel website checklist covers the source content itself.

Where does AI translation fail on hotel text?#

It fails on terms of art, on anything legal, and on anything it decides to complete for you. Plain descriptive prose is where it is strongest.

Board types are the classic trap. Her şey dahil, tam pansiyon, yarım pansiyon and oda kahvaltı map onto all inclusive, full board, half board and bed and breakfast, but the boundaries are hotel-specific: which drinks, which meals, until what hour. Write those boundaries out in each language instead of translating a label. "All inclusive" without the boundaries invites the argument at the pool bar.

Room names carry expectations. Standart, deluxe, aile and çatı katı promise size, layout or a view, and a model may choose a synonym that promises something the room does not have. Cancellation terms fail because models smooth over conditions: free cancellation needs the deadline, the time zone and what counts as the start of the stay, and conditions read awkwardly, so the model quietly simplifies them.

Numbers, dates, currencies and units are the characters that cost money: 1.500,00 TL against 1,500.00, day-first against month-first dates, per person against per room, per night against per stay. Proper names should stay as they are: beach clubs, landmarks, airports, transfer companies and the hotel's own name. A model may translate a name that should not move, or invent an English one.

Register is a silent decision. German and Russian sites often expect formal address, English may not, and a model picks one without telling you. Decide the register once and put it in the glossary. Finally, additions: a model may complete a sentence, add a reassurance or drop a condition it judges redundant. On first-tier pages, read against the source line by line. Guest messages are a separate problem with a separate answer, covered in our article on AI translation for guest messages.

How do you keep five language versions from drifting apart?#

Keep one source text, change it first, and re-translate only what changed. Every other method ends with versions that disagree.

Name one source language, Turkish or English, whichever the marketing team actually writes and approves in, and state it in file names and the content calendar. Never translate from a translation: the second-hand version inherits the first one's errors and adds its own.

Keep a change log per page: the date, what changed, and which languages were updated. Re-translate paragraphs, not whole pages. A model handles a changed paragraph with its surrounding context well; a full retranslation invites silent rewrites of text nobody meant to touch.

Keep a glossary: board types, room names, facility names, the hotel's own name, how you address the guest, currency and date format. Give it to whoever translates and to whoever checks. Show a last-updated date on each language version, so a reader can see that the German page is older than the Turkish one.

When a policy changes, publish it in every language on the same day, or leave the old version up with a clear note. Never let one language promise something the others do not.

Who checks a translated hotel website, and what do they check?#

A fluent reader who did not write the translation checks it against the source, and the reservations team signs off the pages that decide money.

The checker must read the target language natively or near-natively, and must not be the person who produced or prompted the translation. Check in both directions: read the translation and ask what the source says, then read the source and ask what the translation says. One direction alone misses omissions.

The reservations manager signs off cancellation, deposit, child and pet policy, because those are the calls the desk actually takes. Counsel confirms the legal pages. The checker confirms meaning; the lawyer confirms legal effect. These are two different jobs, and neither covers the other.

Read numbers, dates, times, currencies and names aloud. Most errors that cost money sit in those characters. Walk the booking path in that language end to end, including the confirmation email, before the version goes live.

The quarterly spot-check#

Take one page per language at random each quarter, have a native reader compare it with the source, and log the result. A language that fails twice returns to full review. The spot-check exists because drift is quiet: nobody decides to let the German page go stale, it just happens between seasons.

How do search engines find the language versions?#

Give each language its own address and tell the search engine which addresses are versions of the same page. That label is called hreflang, and it is a labelling job, not a ranking promise.

Use one address per language, for example /de/, /ru/, /en/, rather than one address that swaps language by cookie or browser setting. A crawler and a shared link both need a stable address: if the address shows German today and English tomorrow, the search engine cannot index either properly.

Hreflang in plain words: a small label in the page code saying that this page also exists in German at that address. Each version lists all versions, including itself. The labels must point at each other: a one-way label is ignored, so a half-finished setup does nothing.

The text on each address must actually be in that language. Swapping a few labels on the same English text does not create a language version. Keep the language switcher visible to people as well, and let a visitor override an automatic guess based on browser settings. Keep the same page structure across languages, so a link to the German cancellation page lands on the German cancellation page and not on the home page.

Make no ranking promise. Correct labels help a search engine show the right version to the right reader; they do not lift a page above better content.

What to do next#

Take one first-tier page, the cancellation and payment terms, and run the whole method on it once. Translate it from the source into your strongest second language, have a fluent reader check it against the source, have the reservations manager sign it off, publish it on its own address with the language labels, then walk a test booking through it. When that page holds, repeat with the price and what it includes, then the allergen and accessibility notes. Languages come from your own records, pages are ranked by cost of error, one source text keeps the versions together, a named checker catches what the model smooths over, and the labels only help a search engine show the right version to the right reader.