Are Hotels Closing Their Doors to AI Agents?

Consolidated findings from independent research using Claude, ChatGPT and Gemini

About this research

This paper consolidates independent research carried out with three AI tools — Anthropic’s Claude, OpenAI’s ChatGPT and Google’s Gemini — together with supplementary desk research by Pertlink. Each tool was asked the same question about the world’s largest hotel groups; their results are brought together here.

These are findings from independent research, not audits, certifications or statements of any hotel group’s policy. They describe what was observed on specific websites, routes and dates.

We strongly suggest that you test your own websites and public-facing information for accessibility, security and data correctness.

Executive summary

Most of the world’s largest hotel groups have not shut out AI agents. In two independent live tests on 18 October 2026 — one run with Claude, one with ChatGPT — agents reached the checkout page at five or six of each tool’s ten groups. Marriott, IHG, Accor and Choice let both agents through.

Access is not consistent, though. Hyatt stopped one agent with a bot challenge and let the other through to payment. Hilton took one agent to payment and returned a gateway error to the other. Wyndham showed neither agent a rate. The same website can open for one agent and close for another, depending on the agent, its network, its location, and the route it takes.

The Chinese majors are a separate case. Jin Jiang, BTG Homeinns and H World put booking behind mobile-number logins or hand guests to an app. That is channel design rather than an anti-agent policy.

A third tool, Gemini, reported that every group blocked reservations. Its paper describes no test method, dates, or properties, and the two live tests did not reproduce that result. We treat its table as a hypothesis — and as a reminder that AI-generated research needs checking.

Taken together: hotels have not closed their doors to AI agents, but most have not decided how to open them either. Today, access is often decided by bot-management settings rather than by policy.

Background

On 17 October 2026, The Verge published “Websites are closing their doors to AI agents” by Emma Roth. It describes users of agents from Meta, xAI, OpenAI, and Instinct meeting blocks across major websites. Delta is stopping agents from booking tickets, Yelp admits only agents that pay for its data, and JPMorgan Chase and Charles Schwab block agents from accounts. Amazon’s legal dispute with Perplexity continues. Much of the friction, the article notes, comes from anti-bot systems that catch agents alongside scrapers and fraudsters.

Hotels were already in the frame. On 2 October, Skift reported that Meta’s Muse agent failed at the room-and-rate step on Hilton, Marriott and Hyatt websites, while another agent, Instinct, completed searches on all three. Hilton told Skift it accepts bookings from AI agents and has not blocked Muse. On 12 October, Google warned the WTTC Global Summit that hotels could become invisible to AI booking agents if their information is incomplete or inconsistent.

This paper puts The Verge’s question directly to the hotel industry: when an AI agent arrives at a hotel group’s own website, can it read the information and make a reservation?

How the research was done

Each AI tool was asked to examine the websites of the ten largest hotel groups and report whether it could pull information and make a reservation. The approaches differed, and that difference matters when reading the results.

The two live samples share nine groups. Claude’s list included Best Western; ChatGPT’s included BTG Homeinns instead, so eleven groups were examined in all. No reservation was submitted, and no personal or payment details were entered in any test.

Consolidated findings

Seven of the eleven groups let at least one agent reach the checkout page, and four did so in both live tests. Information was the easier test: every group’s public pages were readable in at least one live test, with live rates shown everywhere except H World and Wyndham. BTG Homeinns showed prices, but some were unreadable as text.

“Checkout” means a room and rate were selected and the guest-details or payment page opened. No test went further. Gemini’s column reproduces its paper’s reported results, which could not be verified.

Where the tools agree

• Public information is broadly reachable. Property pages, rooms and, at most, groups, live rates were readable to an agent. Gemini also reports static content as accessible, although it describes live rates as blocked.
• Four groups are open to agents today. Marriott, IHG, Accor and Choice took both live agents to checkout, from different cities and different networks.
• China’s majors gate booking by design. All three papers describe Jin Jiang and H World routing bookings through mobile-number logins, SMS codes, apps, or WeChat; ChatGPT found the same at BTG Homeinns.
• Payment and identity are the last mile. No test completed a booking. Every paper points to card security, verification steps, or authentication as the stage agents have not yet cleared.
• A refused agent goes elsewhere. Claude and Gemini argue that agents turned away by hotel websites will book through intermediaries instead. ChatGPT treats this as plausible but untested, since none of the tests measured guest behavior or revenue.

Where the tools disagree

Hyatt. Claude’s agent, browsing from a home connection in the Philippines, met a Kasada bot challenge and saw only blank pages. ChatGPT’s cloud browser reached the payment page at Andaz Singapore without signing in. A likely explanation is that bot defenses score each visitor differently by network, location, browser, and behavior.

Hilton. Claude reached the payment step at Tempo by Hilton Times Square. ChatGPT received a “502 Bad Gateway” error at Hilton Singapore Orchard, which did not name any bot restrictions. Skift’s earlier tests saw Meta’s Muse fail at Hilton’s room-and-rate step while Instinct succeeded.

Wyndham. Neither agent saw a rate. Claude’s availability request was dropped right after Akamai’s bot-detection sensor reported in. ChatGPT’s Singapore property reported no availability. An agent — and the guest behind it — cannot tell a sold-out hotel from a silent refusal.

Gemini against the live tests. Gemini reported every group as blocked at reservation and named a security vendor for each. Its paper gives no test date, properties, or evidence. Several of its claims conflict with direct observation: both live agents reached Choice’s payment page, which Gemini lists as blocked, and Claude observed Kasada at Hyatt, where Gemini names PerimeterX and AWS. Its strategic framework is useful; its results table should be read as hypothesis, not evidence.

Policy, protection and the new plumbing What hotels publish

Claude’s review of the robots.txt and llms.txt files found that none of the groups whose files it could read named AI agents to block them. Choice and Best Western explicitly welcome named AI agents and crawlers, while IHG, Accor, Wyndham and Choice publish llms.txt guides for AI systems. A wider study of 105,002 hotel websites in seven countries found that only 3.3% set any robots.txt rule against AI crawlers. Formal policy rarely blocks AI; any blocking happens in the bot-management layer.

Industry advisers quoted by PhocusWire have argued for selective policies: admit beneficial AI bots, and block harmful ones.

Why the defenses exist

Travel is a prime target for malicious automation. Imperva’s 2024 Bad Bot Report, as reported by Travel Weekly, found that malicious bots accounted for 44.5% of online travel traffic and that travel accounts for 20.7% of all bot attacks, second only to retail. Bot protection on hotel websites is a reasonable response. The difficulty is that it judges behavior, and a guest’s agent can behave much like a scraper.

Own agents and emerging standards

Some groups are building their own route into AI platforms. Wyndham’s chief executive says the group has placed its own agents inside ChatGPT, Claude and Gemini, and its ChatGPT app hands guests to wyndhamhotels.com to finish booking.

Standards that separate a guest’s agent from a scraper are also arriving. Cloudflare’s “signed agents”, announced in August 2025, let agent platforms sign their requests cryptographically so websites can recognize them. Google’s Agent Payments Protocol (AP2), launched in September 2025, uses signed mandates that set what an agent may buy, with the user approving the final purchase. Gemini’s paper adds the Model Context Protocol as a way for hotels to offer availability and booking to agents directly. None of the tests observed these in use on hotel group websites.

What the findings suggest

Access is variable, not walled. The same website gave different results to different agents on the same day. Neither a guest nor a hotel can predict an agent’s experience without testing it.

“Sold out” and “blocked” can look identical. When an availability search returns nothing, the agent cannot tell which it is. It will report that there are no rooms, and the guest will look elsewhere.

Policy is being set by default. Published files mostly welcome AI or say nothing, while security tools decide in practice. Few groups appear to have connected the two deliberately.

The research itself needs verifying. One of three AI tools produced confident findings that two live tests contradicted. Hotel teams using AI for research should check its conclusions against direct observation.

Considerations for your team

This is a research paper, not a playbook. The right response depends on each group’s AI strategy, distribution economics, and appetite for risk, as well as on how the market develops. Your own team should develop the game plan. These questions may help frame that work:

1. Can a guest’s agent find, price and book on our website today — using several different agents, from several locations?

2. Is our public information complete, current and consistent across rates, taxes, fees, cancellation terms and amenities?

3. Do our bot defenses block what we intend to block, and only that?

4. Do our robots.txt and llms.txt files describe what our systems actually do?

5. Which emerging standards — signed agents, agent payment mandates, direct agent interfaces — belong on our roadmap, and when?

Disclaimer and limitations

Independent research. These are findings from independent research by Pertlink using AI tools. They are not audits or certifications, and they do not state any hotel group’s policy. No hotel group was asked to comment on, or reviewed, this paper.

A point in time. Results reflect specific routes, properties, dates, locations and agents on 18 October 2026. Other agents, regions, brands, properties, or visits may produce different results, and bot defenses change frequently.

No bookings completed. Every test stopped before submission. Payment, card verification, final confirmation, loyalty logins, and mobile apps were not tested.

AI tools can make mistakes. The live observations were made by AI agents and reviewed by Pertlink. Gemini’s reported results could not be verified and, in places, conflict with the live tests.

Different samples. The two live tests used different rankings, cities, and properties. Room-count rankings vary between sources.

Test your own websites

We strongly suggest that every hotel group and hotel tests its own websites and public-facing information — for accessibility to guests and their agents, for security, and for data correctness.

Terence Ronson

Terence Ronson is the Founder and Managing Director of Pertlink Limited, a boutique hospitality technology and AI advisory consultancy headquartered in Hong Kong since 2000, operating across the Philippines and Asia-Pacific. He chairs the AI Education & Training Subcommittee of the HFTP AI Collective and is an inductee of both the HFTP International Hospitality Technology Hall of Fame and the CHTA Hall of Fame.