The short answer
Yes — AI travel planning has gone multilingual, and an English-only GEO strategy now misses most of the conversation. In September 2026, OpenAI’s Signals consumer-usage data confirmed that more than half of active ChatGPT users on consumer plans predominantly use a language other than English, with Spanish, Portuguese and Arabic the most common. The fastest relative growth since a July 2023 baseline has come from Africa and Asia, and in June 2026 the fastest-growing languages with over one million users were Uzbek, Kazakh and Burmese.
Travel demand points the same direction. UN Tourism’s World Tourism Barometer counted 307 million international tourist arrivals worldwide in Q1 2026 — about 6 million more than the same quarter of 2025 — with Europe up roughly 3% in the first half of 2026. International guests are, by definition, guests who plan in their own language.
The mechanics are subtler than “translate your website.” Reporting on ChatGPT Search’s behavior shows the model often runs background retrieval queries in English even when the traveler’s prompt is in Spanish or Arabic, then composes the final answer in the traveler’s language. That means you need two layers: a machine-readable English facts layer that retrieval can find, and a local-language trust layer that the model uses to corroborate and recommend — and that the guest uses to decide whether to book you directly. A travel AI audit gives you the baseline for both.
The user base flipped in 2026
For three years, English-first content strategy was a defensible default: AI assistants were mostly used in English-speaking markets. That assumption is now dead.
OpenAI’s Signals release (reported by Search Engine Journal, September 2026) covers individual ChatGPT plans — Free, Go, Plus and Pro — and shows weekly active users growing on every continent since July 2023, with non-English usage growing fastest in lower-HDI countries. OpenAI also reports that after six months on ChatGPT, people send about 50% more messages per day and attempt twice as many unique tasks. Habitual, multi-purpose use is exactly the pattern that turns an assistant into a trip-planning tool.
Pair that with what the Princeton-led GEO research (KDD 2024) established: generative engines reward sources that carry citations, quotations and statistics, lifting source visibility by up to 40%, while keyword stuffing actively hurts performance. And per Muck Rack’s “What Is AI Reading?” study (December 2025), about 82% of AI citations come from earned media — not brands’ own sites. The models are reading the web the traveler actually reads, in the languages the traveler actually reads.
How AI answers a German question: the two-layer reality
When a traveler in Munich asks an AI planner welches Boutique-Hotel in Lissabon ist ruhig und zentral?, three things happen:
- Retrieval often fires in English. The model’s background searches frequently run in English regardless of the prompt language, as Search Engine Journal documented in its fan-out query reporting. Your English-language facts — schema markup, room descriptions, policies, prices — are what retrieval is most likely to fetch.
- Corroboration happens across languages. The model cross-checks what it found against German-language reviews, German OTA listings, German travel media and the German-language web about your property. If your German-language footprint is thin, contradictory or stale, that weakness surfaces in the answer — or your competitor gets cited instead.
- The answer renders in German. Whatever the model found in English gets rewritten for the German traveler, including the recommendation set.
This is why both layers matter and neither alone is enough. English-only optimization gets you retrieved; local-language trust gets you recommended and booked. The long-standing CSA Research “Can’t Read, Won’t Buy” finding — roughly 72% of consumers spend most of their online time on sites in their own language, and a similar share say they’re more likely to buy with information in their own language — has simply been inherited by the AI layer.
Why machine-translated pages alone fail
Most hotels that “go multilingual” publish machine translations of every page and stop there. Four things break:
- Thin-content risk. Auto-translated pages that duplicate the English source add no new facts for a model to cite, and can be treated as near-duplicates.
- Entity fragmentation. The German page calls the property “Hotel am Meer,” the booking engine says “Hotel Am Meer,” Google Business Profile says “Am Meer Hotel” and the German press writes “Hotel Ameer.” To a retrieval system these may look like four businesses. Entity consistency across every language version is a GEO requirement, not a cosmetic one.
- No local-language third-party proof. With ~82% of AI citations coming from earned media, a property invisible in a language’s review sites and travel press stays invisible in that language’s answers, no matter how many translated pages it publishes.
- Missing hreflang. Without correct hreflang annotations, the models and the classic search layer underneath them cannot reliably determine which version to retrieve, cite or link for which market.
If you haven’t yet hardened the English facts layer itself — schema, consistent naming, crawlable pricing — start with hotel schema markup for AI citations; the multilingual layer inherits every defect the English layer has.
Where the model looks, market by market
| Guest market | How they ask | Where corroboration comes from | What to fix first |
|---|---|---|---|
| German (DACH) | German prompts in ChatGPT, Gemini, Perplexity; Google AI Mode in German | German-language reviews (Google, HolidayCheck), German travel media, DEU OTA listings | German review generation + one strong localized page + hreflang |
| Spanish (Spain + LATAM) | Spanish prompts; fastest-growing AI user base in LATAM | Spanish reviews, LATAM OTAs, Spanish-language destination guides | Spanish GBP description + Spanish review responses |
| French | French prompts, voice assistants | French reviews and travel press, FR metasearch | French FAQ page answering real pre-booking questions |
| Japanese | Japanese prompts; high group-travel planning | Japanese reviews (Google, Rakuten Travel ecosystem), JTB media | Consistent katakana name rendering everywhere |
| Arabic | Arabic prompts; #3 AI language globally | Arabic reviews, regional OTAs, Arabic influencer coverage | RTL-safe localized page + transliteration consistency |
Pick markets from your own booking data, not from general rankings. If 30% of your direct bookings come from Germany and Japan, those are your first two languages — regardless of what the global language charts say.
Tour operators and DMCs get the multiplier effect
For tour operators and DMCs the stakes are higher, because the product is more complex than a room night. AI planners assembling a multi-day itinerary in Portuguese will favor operators whose itineraries, inclusions and safety information exist in Portuguese with consistent entity data — the same trade name, the same tour taxonomy, the same departure details across their own site, partner sites and local listings. A DMC whose German partner pages carry different cancellation terms than its English site is feeding the models contradictory facts, and contradictions are precisely what citation-cross-checking is designed to surface.
The 90-day multilingual GEO plan
Days 1–30 — Measure. Run an AI visibility audit for hotels across your top two guest languages: ask the ten highest-intent planning prompts in each language and record who gets recommended and cited. Inventory your entity footprint per language — name variants, listings, review volumes, hreflang status.
Days 31–60 — Build. For each priority language: publish one genuinely strong localized page (not twenty thin translations), fix hreflang, align your property name rendering, start systematic review generation in that language, and respond to reviews in that language — review responses in the guest’s language are themselves citable content.
Days 61–90 — Earn and convert. Pursue earned media in the target language: destination-marketing organization listings, local travel press, partner cross-linking. Then wire conversion: the localized page must carry a direct-booking path with language-appropriate reassurance (cancellation terms, payment options, contact in that language). For a structured approach across markets, a travel GEO agency engagement typically starts exactly here, and pricing scales with the number of language markets you activate.
FAQ
Do I need to translate my entire website for AI travel planners? No. One strong, fact-dense localized page per priority market, correct hreflang, consistent entity data and local-language reviews beat a full machine-translated site every time. Depth beats coverage in GEO.
Does ChatGPT answer travel questions in the traveler’s language? Yes — and per OpenAI’s September 2026 Signals data, most consumer users now ask in a non-English language. The answer renders in the prompt’s language even when background retrieval runs in English, which is why your English facts layer and local-language trust layer both matter.
Is hreflang still relevant when AI assistants do the retrieving? Yes. Hreflang disambiguates your language versions for the retrieval stack that generative engines sit on top of, and for the traditional search results that still send the clicks that convert AI referrals into direct bookings.
Which languages should a hotel prioritize for GEO? Rank by your direct-booking and inquiry data: source markets sending the most revenue come first. Globally, Spanish, Portuguese and Arabic lead AI assistant usage — but a Kyoto ryokan should rank Japanese, Korean and Mandarin before any of them.
How do I know multilingual GEO is working? Track three things: recommendation share on your priority prompts per language, citation sources by language, and direct-booking or inquiry volume by guest language. Traffic alone is not the metric — recommendation presence is.
Where this fits
Multilingual GEO is not a separate discipline from your AI visibility program — it is the same program, extended to the languages your guests actually think in. Audit both layers, fix the entity layer once, then expand one language market at a time. The properties that do this in 2026 will be default answers in every language their guests speak.