Illustration: AI-generated for EA Hospitality Pulse — synthetic image, not a photograph

Nobody at your lodge decided to block the AI agent. Something in your stack did.

By Onyango George (OG)

On 18 October an AI agent working for a traveller could not see a rate at a Wyndham hotel. On 19 October it reached the guest details page. Nothing in Wyndham's published policy changed overnight, and that is the point: access to hotel inventory is now decided by bot-management software rather than by anyone's commercial judgement. East African operators have never looked at what their own stack does.

On Sunday 18 October an AI agent acting for a traveller went looking for a room in a Wyndham hotel and never got as far as a price. On Monday 19 October the same agent, pointed at the same chain, reached the guest details page. Nothing in Wyndham's published terms changed in the intervening twenty four hours. No executive signed off a new policy on AI agents over the weekend. What changed was a score, assigned in milliseconds by a piece of bot-management software, about whether the visitor looked like a person.

The tests were run by Terence Ronson, who has spent three decades on hotel technology and runs the Hong Kong consultancy Pertlink. He pointed two commercial AI agents at the websites of the world's ten largest hotel groups on 18 and 19 October, instructed them to behave like a guest, and stopped them short of entering payment details. Readers of his first write-up asked for the evidence, so on 20 October he published the screenshots and the HTTP status codes. Marriott, InterContinental, Accor and Choice let both agents through to checkout on every attempt. Hyatt blocked one agent with an HTTP 429 at a challenge page served by the bot-management vendor Kasada, while the other agent reached the payment screen at Andaz Singapore. Hilton let one agent reach payment three times and handed the other a 502 error. The large Chinese groups, Jin Jiang, H World and BTG Homeinns, sat behind mobile-number logins, SMS codes and apps, which Ronson reads as channel design rather than an anti-agent stance.

"Most of the world's largest hotel groups have not shut out AI agents," he concluded. The more useful sentence is the next one: "Access is often decided by bot-management settings rather than by policy." That is not a story about hoteliers taking a view on artificial intelligence. It is a story about hoteliers not taking one, and discovering that a vendor default took it for them.

The written record makes the gap plain. On 19 October the agents read the robots.txt files of seven of those groups. Not one names an AI agent in order to block it. Choice and Best Western name AI agents only in order to allow them. Marriott, Hilton, InterContinental, Accor and Wyndham all disallow their search and booking paths to every automated visitor, which on paper is a flat refusal, and the agents walked through anyway. The document that states the policy and the system that enforces one are, in most of the industry, not connected to each other.

That disconnection is not confined to the majors. Nicolas Sitter, who researches how hotel websites are read by machines, parsed the robots.txt files of 105,002 reachable hotel domains across France, Italy, Spain, the Netherlands, the United States, the United Kingdom and Germany. The headline is a near-vacuum: 96.7 per cent carry no AI-specific blocking rule at all, 3.3 per cent block at least one AI crawler, and 0.9 per cent block all of them. "The industry is wide open," he writes. GPTBot is the most-blocked single crawler, at 2.9 per cent of all sites. Where blocking does appear it is rarely a hotelier's choice: France has the highest national rate at 7.5 per cent, and most of that is one cooperative, Logis, which made a single decision for roughly 950 member properties. Strip Logis out and France falls to 2.1 per cent. Sitter's own summary is the line East African owners should sit with: "Most AI blocking is a chain or CMS decision, not an individual hotel decision."

Nothing in either dataset is African. Sitter's seven countries are European and American, and Ronson's eleven groups own almost nothing in Kenya, Uganda, Tanzania or Rwanda beyond a handful of city hotels. So the numbers in this piece do not describe East Africa, and anyone who tells you what percentage of Kenyan lodges block AI agents is inventing it. The measurement does not exist. The mechanism, though, travels a great deal better than the measurement, because the mechanism is a set of vendor defaults, and East African properties run on more vendor defaults than most.

Consider who actually decides whether an agent can see a rate at a camp in the Mara or a lodge on the Nyungwe escarpment. There is the content management system and the company hosting it. There is the content delivery network sitting in front of the site, filtering traffic according to rules written by its own security team. There is the booking engine vendor, whose availability calendar and rate display are loaded into the page from a different domain, often inside a frame, often only after a script has run. Three parties, at least, hold the gate. None of them is the lodge, and none of them has ever been asked what it thinks about agentic travel.

Cloudflare, whose network fronts a very large share of small hospitality websites, changed its defaults on 15 September. It is worth being precise about what it did, because the change has been reported more dramatically than it deserves. Cloudflare now sorts crawlers into Search, Training and Agent categories, and from that date new domains joining the network block Training and Agent bots by default on pages that display advertising, while Search stays allowed. A lodge website carrying no ads is not caught by that clause. Two other provisions bite harder. Non-verified bots remain blocked by default, and an agent browsing on a traveller's behalf without a verified cryptographic signature is, to the filter, a non-verified bot. And from 15 September, crawlers that do several jobs at once are governed by all of their behaviours, so a site that had earlier ticked a box to block AI training, and did not opt out before the deadline, also began refusing crawlers that index for search, Googlebot and Bingbot and Applebot among them. A lodge that wanted to keep its photographs out of a training corpus may have quietly removed itself from search results instead.

The second exposure is worse, and almost nobody in the region is looking at it. Before an agent can be blocked at the door it has to know the door exists, and models learn that the hard way, from crawled text. Sitter checked 108,109 hotel domains against the May 2026 snapshot of Common Crawl, the open web archive that feeds a great deal of model training, and found 39.4 per cent of them absent entirely: not a single page captured. Independents fared better than chains, at 61 per cent captured against 45.9 per cent. What predicted absence was not blocking but obscurity, measured as the number and quality of inbound links. "Being in Common Crawl is necessary but not sufficient for a model to surface you," he writes.

Map that onto the way East African beds have been sold for forty years. Demand arrives through tour operators, ground handlers, safari marketplaces and specialist agents. The itinerary page, the review, the glossy gallery and the inbound link all accrue to the intermediary. The lodge's own domain is frequently a brochure that nobody links to, serving its rates from a third party's script. That is a plausible recipe for a property that is commercially successful and digitally invisible, and it is an argument rather than a finding, because the study that would settle it has not been run in this region. Someone should run it.

None of this removes the commission. This publication argued in September that the agent era moves the shopfront rather than the toll booth, and that remains the case: when an agent books a room it is still usually buying from an intermediary that keeps its margin. Access failure is a separate and earlier problem. It costs you the booking before the price is ever discussed, and it costs it to a competitor whose content delivery network happened to be configured differently.

The honest counterweight is that very few bookings currently ride on any of this. Expedia Group's survey of more than 5,700 adults in the United States, the United Kingdom and India, conducted by YouGov in March and published on 15 April, found 53 per cent comfortable letting AI suggest options, 8 per cent comfortable booking through an AI platform, 68 per cent still preferring to book with a trusted brand even where an AI option existed, and 66 per cent unwilling to let an assistant buy anything on their behalf. "Travelers don't have a technology problem with AI. They have a trust problem," said Xavi Amatriain, the company's chief AI and data officer. On that evidence an East African operator who ignores agent access for another year will probably lose no measurable revenue.

What makes the work worth doing anyway is the asymmetry. Checking costs a morning. Being the one lodge of six that an agent could not read costs the entire booking, silently, with no enquiry in the inbox to tell you it happened. The people building the agents are blunt about the direction. Meta, pressed this month on retailers and airlines whose sites its agent cannot navigate, said that "turning away a personal agent means turning away the customer behind it", and that the fix starts with clear norms. Those norms do not yet exist. Meta, Walmart, Stripe and several others are drafting an open standard for how agents identify themselves to businesses, and until something like it is settled, every site is guessing and every filter is guessing back. Cloudflare, asked by TechCrunch on 6 October how often personal agents are being blocked, said it had no specific data.

There is an irony in the Kenyan position worth naming. In February, at the World Governments Summit in Dubai, Rebecca Miano told a panel that Kenya was "leveraging AI and smart data to personalise the visitor experience" and that the Global South was "no longer a passive player". The state is marketing an AI-mediated destination with some energy. Nobody has yet audited whether the destination's inventory is machine-readable.

The audit is not difficult. Read your own robots.txt and check whether it names an AI agent in order to block one, which it almost certainly does not, and that is the correct answer. Ask whoever runs your content delivery network whether unverified bots are challenged, and what happens to an agent that arrives from a residential network in Frankfurt. Load your own booking page with scripts disabled and see whether a rate is visible without them. Search for your domain in Common Crawl and find out whether a model has ever read your own words about your own property. Then count the links pointing at your site rather than at the operator who sells you, because that number, more than any robots.txt rule, decides whether you exist to a machine.

Four of the eleven groups Ronson tested let both agents through on both days. None of them did it on purpose either. They were simply not blocking, which is a position, but a fragile one, held by default and revocable by a vendor's next release note. The properties that are reachable in 2027 will be the ones that stopped holding it by accident.