AI in the in-destination layer: museums, heritage, tours, restaurants, attractions and car hire (2022–2026)
The booking layer asks who captures the transaction. The layer that starts when the visitor arrives asks who shapes discovery and the day — and most of what is said about AI in travel is about the first layer.
The one-page summary condenses the paper to five findings, adds a short list of questions to put to an AI vendor, and closes with the product direction the author draws from them.
In flights and hotels, the central AI question is who captures the transaction. In museums, tours, restaurants and attractions it is who shapes discovery and the on-site experience. That distinction is the paper's organising idea, and the evidence gathered through summer 2026 is consistent with it.
What changed over those three months is who is competing to shape discovery. Google, OpenAI, Anthropic and the experience marketplaces have all moved into in-destination discovery, and the destination marketing organisation has become one of the most visibly contested segments in the evidence reviewed. The paper adds a section on the destination layer and a focus section on the agentic harnesses now running in production.
It also sets out a working definition of disintermediation — the loss of discovery control, transaction ownership, customer relationship and data, attribution, or margin — because continued physical fulfilment is not, by itself, evidence that a venue is unaffected.
LLM-referred traffic is well under 1% of room nights at the largest OTA. In the same summer, 88% of AI-using travellers in Germany and 92% in France said AI shaped an actual trip decision. Different populations and denominators, not a funnel — but referral attribution plainly captures only part of the influence.
Every production harness described in public is a booking-layer system: specialists that search inventory, mandates that authorise payment. None is described as re-planning the day as it unfolds, using hours, queues, weather and a curated local corpus — which is where travellers actually report using AI in destination.
The industry has converged on the Model Context Protocol, so connecting an agent to inventory is no longer the hard part. Commercial access is, and it is gated everywhere: terms unpublished, servers inside enterprise contracts, one major platform with no open public API.
Europeans report using AI more while trusting it less than friends, family and human advisers. Of surveyed US DMO leaders, 22% offer an AI trip planner while 63% of travellers would use one. The scarce asset is not AI capability; it is being the source the assistant draws on.
The Digital Omnibus deferred the AI Act's high-risk rules to 2027 and 2028 but left Article 50 on its original date. An EU-facing chatbot must make clear the user is dealing with an AI, and generative output needs machine-readable marking. Fines reach €15m or 3% of turnover. Easy to meet — but the contract should say who carries it.
Every statistic, date, benchmark result and legal proposition carries a numbered citation, and sources are labelled by evidence type — peer-reviewed, preprint, law and guidance, earnings disclosure, product record, vendor release, vendor survey, trade press or general news. Vendor material is described as verifying what a company said or launched, not adoption, accuracy or return.
Qualitative judgements are not presented as measurements, indicators with unlike denominators are not plotted on a shared axis, and interpretation is distinguished from disclosed fact. The September draft was submitted to an independent evidence audit before release; six statements found to be misstated or overstated were corrected, and the corrections are logged in the paper rather than quietly absorbed.
Descriptive in intent; interpretation is the author's unless attributed. Legal references are informational analysis, not legal advice.