Are Hotels Closing Their Doors to AI Agents? Part Two

Part Two: Showing our work — the supporting evidence behind our findings, in response to reader feedback

Comparison covers the nine groups tested live by both Claude and ChatGPT, across three runs: Claude on 18 and 19 October, ChatGPT on 18 October. Best Western and BTG Homeinns were each tested by one tool and are reported separately.

About Part Two

Part One, published on 18 October, consolidated independent research by three AI tools — Claude, ChatGPT and Gemini — into whether the world’s largest hotel groups let AI agents in. A reader’s response asked a fair question: where is the evidence?

Part Two answers it. It sets out the screenshots, HTTP status codes and repeat tests behind every live result Claude reported, re-tests the results on which the tools disagreed, and corrects Part One where the evidence requires.

These remain findings from independent research. We strongly suggest that you test your own websites and public-facing information for accessibility, security and data correctness.

Introduction

The evidence supports Part One’s broad picture, with sharper edges. Repeating the tests on 19 October, Marriott, IHG, Accor and Choice took an AI agent to checkout for the third time in three runs. Jin Jiang and H World still put booking behind a mobile-number login or an app.

Two results changed. Wyndham, which refused Claude’s availability search on 18 October, took the same agent to its guest-details page a day later. When Claude re-ran the routes where ChatGPT had reported different outcomes, ChatGPT’s Hilton error and Wyndham “no availability” result did not recur. Hyatt, meanwhile, blocked Claude on the very property where ChatGPT had reached payment.

The clearest finding is about who decides. No robots.txt file we could read names an AI agent in order to block it. Access is decided by bot-management settings: Hyatt’s block came from Kasada’s challenge software, yet Choice runs the same software and let both agents through.

Part One said most of the largest groups had not shut AI agents out. The evidence shows no deliberate closure — but it does not show the industry is open either. Two agents tested over two days, with access changing by agent and by day, cannot settle that.

What the reader asked, and what we did

How the evidence was gathered

Claude repeated its ten live routes between 07:40 and 07:58 Manila time on 19 October 2026. It used the same browser, home connection in the Philippines, properties, and stay (one adult, 18–19 November) as Part One. At each stop point, it recorded:

A screenshot of the page.

The HTTP status code of the page, read from the browser’s own timing record.

Bot-management and CAPTCHA scripts loaded on the page — Akamai, Kasada, Imperva, reCAPTCHA, hCaptcha, and others.

The form fields present, confirming a guest-details or payment page.

Claude also fetched each group’s robots.txt file directly, and re-ran ChatGPT’s three disputed routes on ChatGPT’s own properties and dates (17–18 November). The stop rule was unchanged: no personal or payment details were entered, and no reservation was submitted.

The evidence, run by run

Four of the nine groups reached checkout in every run; three changed between runs; two gate booking behind a login or an app.

“Checkout” means a room and rate were selected and the guest-details or payment page opened. No test went further. Status codes are for the page where each run stopped; full records are in the Appendix. ChatGPT’s results are as reported in its paper.

Tested by one tool only

Same property, different agent

On ChatGPT’s own properties and dates, Claude’s results differed in all three disputed cases. The differences follow the agent and the day, not the property.

Who decides access

Several observations point to bot-management settings, rather than published policy, deciding whether an agent gets in:

Hyatt. Claude’s browser received HTTP 429 with Kasada’s challenge script on every page tried — the New York search, the homepage, robots.txt and Andaz Singapore — on both days. ChatGPT’s cloud browser reached payment at Andaz Singapore on 18 October.

Choice runs the same vendor. Kasada’s script loads on Choice’s booking page alongside Akamai’s, and both agents passed on every run. The vendor alone does not decide the outcome; its settings do.

Hilton runs Akamai and let Claude through three times, in New York and Singapore. ChatGPT’s 502 error named no bot restriction and did not recur on the same route a day later.

Wyndham changed overnight. On 18 October, Claude’s availability request was dropped mid-connection, straight after Akamai’s sensor exchange, and ChatGPT saw no availability in Singapore. On 19 October, the same request returned 71 New York hotels, and Singapore showed rates.

Published policy points elsewhere. Claude read each group’s robots.txt file directly on 19 October (Appendix C). None of the seven files it could read names an AI agent in order to block it; Choice and Best Western name AI agents only to allow them. Marriott, Hilton, IHG, Accor and Wyndham disallow search or booking pages for all automated visitors — rules our agents passed in practice. Hyatt’s file returned the same 429 challenge, and Jin Jiang’s WeHotel site and H World have none.

Part One also cited a separate study by Nicolas Sitter, which found 3.3% of 105,002 hotel websites in seven countries set any robots.txt rule against AI crawlers. That study covers a different, broader population, mostly individual hotels. It is context, not a finding about the groups tested here.

We cannot see inside these systems. The observations are consistent with bot scoring that varies by agent, network, location, and time, but they do not prove it. A temporary fault or a genuine inventory gap could explain individual results.

What changes from Part One

Where the evidence still stops

ChatGPT’s results are text-only. Its paper reports observations and one status code. Where Claude re-ran the same route, the result is documented; ChatGPT’s other results rest on its own account.

Gemini’s results remain unverified. Its paper describes no method or evidence, and several of its claims conflict with the live tests.

Two agents, two days. Bot scoring can vary by agent, network, location, and time. Other agents may meet different results.

No bookings were completed. CAPTCHA scripts load on Best Western’s and Wyndham’s booking pages but were not triggered before our stop point. Payment and confirmation were not tested.

Vendor detection has limits. Vendors were identified from page scripts. During the re-run, Claude flagged Kasada on Hilton in error — an unrelated advertising script matched — and corrected it before publication.

Right of reply

No hotel group was asked to review either part of this report before publication. The results describe what our agents observed on specific dates and routes; they are not statements of any group’s intentions or policy. Pertlink welcomes responses from the groups named and will add any it receives to future versions.

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. Test with more than one agent, from more than one location, and repeat the tests over time.

The intelligence may be artificial. But the experience is human.

References

[1]        Ronson, T. / Pertlink (2026, 18 October). Are Hotels Closing Their Doors to AI Agents? Consolidated findings from independent research using Claude, ChatGPT and Gemini (Part One).

[2]       Roth, E. (2026, 17 October). Websites are closing their doors to AI agents. The Verge.

[3]       Claude (Anthropic) for Pertlink (2026, 18 October). Are Hotels Closing Their Doors to AI Agents? A live test of the world’s ten largest hotel groups.

[4]      Claude (Anthropic) for Pertlink (2026, 19 October). Verification re-run and evidence pack: screenshots, status log, and robots.txt reads.

[5]      ChatGPT (OpenAI) for Pertlink (2026, 18 October). Hotel Website AI Access Paper.

[6]      Gemini (Google) for Pertlink (2026). Viewpoint: The Agentic Hospitality Dilemma, version 2.

[7]       Sitter, N. (2026, 20 March; updated 24 June). Hotel robots.txt and AI blocking study 2026.

Appendix: supporting evidence

Claude’s re-run took place between 07:40 and 07:58 Manila time (UTC+8) on 19 October 2026, using the browser in the Claude desktop app from a home connection in the Philippines. The full evidence pack — screenshots at original resolution, the status log, and robots.txt reads as CSV files — accompanies Part Two.

A. Stop-point records, 19 October re-run

* Hilton’s booking steps load within one page; the status shown is for the initial page load. Status codes were read from the browser’s navigation timing record for each page.

B. 18 October records cited in this paper

C. robots.txt reads, 19 October

D. Screenshots

Each image shows the page where the run stopped, or the evidence of a block. Cookie and language banners were left unanswered or declined, and appear in some images.

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.