I read reports from Trivago, Expedia, Booking.com. All of them are saying that their traffic from AI LLMs is less than 1%. And now there is a narrative in the market that LLMs will disintermediate or personal AI will disintermediate Booking or Expedia, places like this. First of all, look, AI has been there for years now and still it has contributed to only less than 1% of traffic. So are people really going to AI to plan the trip or maybe they are just using it to do some organization? I mean the filters Booking or Agoda has already, the interface they have, it is already good enough to find hotels. I am not sure what will AI achieve. Is it so difficult of a problem that AI will solve this problem? I don't know. Please elaborate on this.

The skepticism is backed up by both operational data and basic transaction economics. Booking Holdings confirmed in recent earnings reports that traffic and room nights driven by LLM referrals remain under 1%, and survey data across major travel platforms consistently reveals an AI trust gap: while over half of travelers are open to using generative AI to brainstorm trip ideas, fewer than 10% are willing to let an AI transact on their behalf.

The narrative that a conversational chatbot will easily disintermediate platforms like Booking.com, Agoda, or Expedia fundamentally misunderstands what problem OTAs (Online Travel Agencies) actually solve, why consumers browse the way they do, and where the real moat lies.

1. Relational Filters Beat Natural Language for High-Volume Search

The user experience argument is the most overlooked flaw in the "chat replaces search" hype: for structured inventory, tabular visual browsing is cognitively superior to reading text.

• Speed of visual scanning: A human can open an Agoda or Booking map/grid view, apply four filters (price range, 8.5+ rating, breakfast included, free cancellation), and visually scan 20 options—evaluating room photos, pool aesthetics, bed types, and neighborhood density—in 30 to 45 seconds.
• The conversational tax: Asking a chatbot to suggest hotels returns 3 to 5 text descriptions. Evaluating them requires reading paragraphs, asking follow-up questions, and clicking external links anyway to see what the bathroom looks like. Chat is a low-bandwidth interface for visual decisions.
• Deterministic filtering is a solved problem: Relational databases handle boolean queries (WHERE price <= 150 AND stars >= 4 AND rating >= 8.5) instantly, accurately, and at near-zero compute cost. Running a 70-billion-parameter neural network just to filter by price and star rating is computationally wasteful and prone to hallucinating out-of-date prices or phantom amenities.

2. What Problem AI Actually Solves (And What It Doesn't)

Where traditional filters fall short is unstructured, multi-variable semantic context. Relational databases are brittle when search queries stop looking like database columns:

• Nuanced, qualitative constraints: Filters cannot easily answer: "Find me a boutique hotel within a 10-minute walk of specialty coffee shops, with good natural desk light for remote work, but tucked away on a quiet side alley away from nightlife noise."
• Review synthesis at scale: Travelers often waste hours reading through hundreds of reviews to answer hyper-specific questions: "Is the Wi-Fi stable enough for video calls on the 4th floor?" or "Is the street construction outside still going on?" LLMs excel at synthesizing thousands of unstructured sentiment signals into a single verdict.
• Complex, multi-city logistics: Connecting a multi-leg itinerary (train schedules, check-in windows, luggage drop-offs, and flight arrivals) across disparate services into a coherent schedule is genuinely tedious.

AI does not replace the filter; it replaces the research manual labor that happens before applying the filter. Travelers are happily using AI as an interactive guidebook and itinerary organizer, but once they identify the neighborhood or the exact property name, they switch to an OTA or the hotel’s direct app.

3. The "AI Trust Gap" and High-Stakes Transactions

Booking travel is fundamentally different from searching for code snippets, recipe ideas, or web articles:

When a traveler books a non-refundable, $300-a-night hotel room across the globe, they require certainty:

1. Financial recourse: If the hotel claims it has no record of the booking at midnight, who is accountable?
2. Cancellation and modification: A consumer trusts Booking’s or Expedia’s one-click "Cancel Booking" button and clear refund timeline far more than trusting an autonomous agent to negotiate a refund.

4. The Real Moat: Plumbing, Merchant of Record, and Fragmented Supply

The market narrative of "AI disintermediating OTAs" assumes the hardest part of travel is the user interface. It is not. The hardest part of travel is the unglamorous, messy back-end infrastructure.

[ Independent Hotel / Boutique Property ]
│
[ Channel Manager / PMS ]  (SiteMinder, Cloudbeds, Opera)
│
[ OTA Aggregator ]       (Booking / Expedia / Agoda)
┌─────────┴─────────┐
▼                   ▼
[ Direct Apps ]     [ LLM / Agent ]
(Consumer Trust)    (Front-end router)

1. Fragmented Supply & Direct Contracting: Booking.com and Expedia have spent 25+ years deploying thousands of market managers worldwide to sign contracts with millions of independent hotels, boutique villas, and bed & breakfasts. These hotels do not have standardized, public, modern APIs that an AI bot can simply tap into. They rely on legacy Property Management Systems (PMS) and channel managers (e.g., SiteMinder, Cloudbeds, Amadeus).
2. Merchant of Record (MoR): OTAs handle local sales taxes, multi-currency processing, fraud prevention, chargeback handling, and localized payment rails (Alipay, PromptPay, iDEAL, Pix). An AI company like OpenAI or Anthropic has no interest in operating a global 24/7 call center or acting as a travel merchant of record liable for room overselling.
3. Wholesale & Loyalty Rates: OTAs negotiate discounted closed-user-group rates (such as Booking’s Genius discounts or Agoda’s VIP tiers) in exchange for massive guaranteed booking volume. A standalone personal AI searching the open web cannot unlock those closed-door negotiated margins without passing through the OTA pipe anyway.

5. Absorption, Not Disintermediation

Rather than being replaced, the major travel platforms are simply absorbing AI:

• Backend Agents via OTA APIs: If autonomous personal agents eventually do book trips for users, they will not bypass OTAs; they will execute transactions through them. Booking and Expedia have already built plugins and API integrations for platforms like ChatGPT. The OTA remains the clearinghouse, inventory provider, and merchant of record; the AI is merely a new referral front-end (replacing traditional Google search ad clicks).
• Native In-App AI: OTAs are embedding conversational search directly inside their apps. This gives travelers the best of both worlds: natural language querying ("find family-friendly villas near the beach with a kitchen") directly layered on top of live inventory, member pricing, and 1-click checkout.

The current sub-1% traffic metric is not a temporary blip waiting for better prompt engineering; it reflects a structural reality. Travel planning is open and exploratory (where AI excels), but travel booking is transactional, visual, and risk-averse (where established platforms dominate).

Want to look into how the backend plumbing (PMS & channel managers) actually blocks AI agents?