AI Visibility in Hospitality 2025

The new reality: AI is replacing search

For two decades, finding a hotel meant a keyword, a list of blue links, and a booking. AI-powered discovery has no blue links — the assistant either recommends your property or it does not. To map this new landscape, Nokumo Research ran one study with two components. This Nokumo Research study has two components: (A) an analysis of AI assistant responses to 450 hospitality prompts and (B) a technical audit of 1,337 accommodation websites. The components are linked: 76 audited websites were cited by at least one model. Component A tested 450 prompts across four AI models (GPT-5, Gemini 2.5, Perplexity and Mistral), yielding 3,600 responses, 20,370 URL citations and 13,859 brand mentions; component B audited 1,337 accommodation websites across 30,935 pages in five markets — Germany, Austria, Switzerland, Slovenia and Croatia. Data for both components was collected between May and August 2025.

Booking.com — the undisputed champion

One domain towers over all others. Booking.com appears in 95.3% of every query tested and takes 14.5% of all citations — one in seven. The top three domains (Booking.com, TripAdvisor, Airbnb) account for 19.5% of citations, and only 86 of 4,043 domains are cited across all five countries. For every independent-hotel URL an AI cites, it cites roughly two OTA URLs.

Four AI models, four personalities

The models differ enough to be statistically distinct (χ² = 494.63, p < 0.001). GPT-5 is the most generous citer at 10.6 URLs per response and gives the best direct-booking outcomes. Perplexity is the most predictable and leans on review and UGC sources. Gemini 2.5 is the most OTA-heavy — significant given its place inside Google Search. Mistral returns zero URL citations yet still names 2.5 brands per answer.

What travelers ask, who AI recommends

Intent is destiny. Luxury, wellness and business queries surface direct hotel sites most often; budget-city, family and event queries are captured by OTAs. Event and festival queries are the OTA trap — 39.4% OTA dependency, the highest of any intent — so hotels near a venue should publish event-specific pages with direct booking CTAs.

The language effect

Every query was asked in English, German and Croatian. Language changes the answer. Croatian-language queries produce the highest OTA dependency in almost every market — up to ten percentage points higher — likely because Croatian-language training data is thinner on local providers, so models fall back to globally indexed OTAs. For DACH properties, German-language queries deliver the best direct-booking outcomes.

1,337 websites under the microscope

The technical audit found an industry operating at under 40% of its digital potential, with an average score of 38.1 / 100. 77.1% of properties have no booking engine, only 39.2% use any schema markup, and just 7% implement Hotel/LodgingBusiness schema. Most decisively, only 5.7% of audited properties were detected in any AI model response at all.

What AI-cited properties do differently

Of 1,337 properties, only 76 (5.7%) were cited by any model — and they are not the flashiest sites, they are the clearest. URL quality is the single biggest differentiator (Cohen's d = 0.60), followed by trust signals (0.50) and content quality (0.36). Schema, while worth doing, shows the smallest measured effect. A well-written 10-page site beats a flashy 50-page one.

Content strategy: what AI wants vs. what exists

The corpus is dominated by basic property information (61.1% of pages) while the categories AI actually queries on are missing. Zero percent of audited sites had a dedicated room-listing, dining or amenities page. The biggest content gap is budget and value — 74.8% of properties have none — followed by luxury positioning and events. Even a basic five-page site with entity-rich content outperforms 60% of the corpus.

The playbook: what to do now

The actions that move the needle, in order. Today: complete and maintain your Booking.com and TripAdvisor listings, and list on your national and regional DMO. This week: sharpen a specific, factual positioning and build intent-specific pages for luxury, wellness or business travel. This month: add Hotel/LodgingBusiness schema and make sure robots.txt allows GPTBot, Google-Extended, ClaudeBot and PerplexityBot. Ongoing: build external authority through editorial coverage, and re-test 10–20 representative prompts every quarter.

Relationship to the preliminary release

Preliminary findings from this study were presented at VRM Days Croatia (Zagreb, 22 October 2025) and covered by the press in November 2025. The figure of 88% reported in the press in November 2025 came from the preliminary release. It was based on the uncleaned preview sample of about 1,500 websites, where roughly 12% had lodging-related structured data. It was a media teaser and is not part of this report's analysis. This report's schema figures (39.2% of sites use any schema; 7% use Hotel or LodgingBusiness schema) come from the cleaned sample of 1,337 websites and are not comparable with the preview figure. This report supersedes the preliminary figures.

Coverage and presentations

Press coverage of this study, from the preliminary findings presented at VRM Days Croatia (22 October 2025) to the full report (23 March 2026).

• HrTurizam · 2025-11-03 · Preliminary findings (Nov 2025)

  AI ne vidi vaš hotel: Nokumovo istraživanje otkriva da 88% smještajnih webova u regiji uopće nije spremno za umjetnu inteligenciju

• Večernji list · 2025-11-09 · Preliminary findings (Nov 2025)

  Čak 88% hotela i drugog smještaja nevidljivo umjetnoj inteligenciji

• fvw | TravelTalk · 2026-03-26 · Full report (Mar 2026)

  Studie zur KI-Visibilität: Wie Ferienhäuser sichtbar bleiben bei ChatGPT & Co

Full press archive

Methodology & AI disclosure

This Nokumo Research study has two components: (A) an analysis of AI assistant responses to 450 hospitality prompts and (B) a technical audit of 1,337 accommodation websites. The components are linked: 76 audited websites were cited by at least one model. Data collection: May-August 2025 (website audit and AI prompts; GPT-5 was added at the end of collection). Deep analysis and report writing: after fieldwork, to February 2026. First public release: preliminary findings, 22 October 2025; full report, 23 March 2026. 450 prompts (150 base prompts in each of 3 languages: English, German, Croatian) were run on 4 AI models, 2 runs per prompt per model: 3,600 responses (20,370 URL citations, 13,859 brand mentions). About 1,500 websites were crawled. About 160 were excluded as noise or unreliable data. 1,337 websites (30,935 pages) were analysed. Research design, analysis and conclusions are human-generated and independently verified; AI tools were used only for editorial assistance and source retrieval. © 2026 Nokumo Services d.o.o. (Infranet Group). Free to share with attribution (CC BY 4.0, with logo).