What the panel said
At the Extraordinary Travel Festival in Bangkok, Travel Massive's Travel Tech Panel, moderated by Travel Massive founder Ian Cumming, covered how some of the world's most travelled people plan and book trips. Several points matter directly to tour and activity operators.
Agents are booking, not just suggesting. Joel Oleson, an enterprise architect at Sony, gave his agent his constraints and let it find which balloon companies flew over Bagan on which days, then fill in the forms and book. He has used personal agents for about a year.
Agents get it wrong when information is unclear. Using Gemini, Oleson asked an agent to book a missing San Diego to Ottawa leg around his existing flights. It looked right, but it left him two days in Ottawa, and he had to change the flight manually. An agent acts on what it can read. Vague dates, schedules or availability lead to wrong bookings, and the operator often takes the blame.
Customers are already arriving via AI. Rachel Weiss, Digital Marketing Manager at Lost Plate food tours, said the company kept hearing "I used AI" when asking customers how they found them, so it added AI as an option in its post-booking survey.
Planning with AI is common; agents are still early. A show of hands found almost everyone in the room used AI to plan or book, while far fewer used agents. That is the window for operators to prepare.
Stale information loses. Weiss noted that Google and Tripadvisor results in China are often out of date, some still from 2019. AI tools draw on the same kind of sources, so an operator's own site needs to be the clearest, most current record of what it offers.
For the wider shift from search to agents, see NJoy's earlier piece on Nomads.com's AI agent API and MCP.
Agents of this kind are already in everyday use. Grok Bot, for example, can open websites in its own browser, fill in forms and book flights, stays, restaurants and rides for the person it works for. A follow-up NJoy test puts an agent through the booking flow on each operator site in this check.
The check: can an agent read and book these tours?
On 28 October 2026, NJoy checked one representative tour or product page from each of 14 tour and activity operators across Thailand, Vietnam, Cambodia, Myanmar and Japan, a mix of small independents and established brands. Operators are anonymised.
| Operator | Type, country | Price on tour page | Real dates or availability | Booking step in raw HTML | Booking needs JavaScript | Product/Offer/TouristTrip schema | AI search crawlers blocked* | llms.txt |
|---|---|---|---|---|---|---|---|---|
| A | Balloon flights, Myanmar | Yes | No | No | No | No | None | No |
| B | Food tours, multi-country | Yes | Yes | Yes | No | No | None | No |
| C | Food tours, Thailand | Yes | Yes | No | No | No | None | No |
| D | Cycling tours, multi-country | Yes | No | No | No | Yes | None | No |
| E | Food tours, Vietnam | Yes | No | No | No | No | None | No |
| F | Food and motorbike tours, Vietnam | Yes | No | No | Yes | Yes | None | No |
| G | Tours marketplace, Japan | No | No | No | Yes | Yes | None | No |
| H | Tour operator, Japan | Yes | Yes | No | Yes | Yes | None | No |
| I | Local guides marketplace, Thailand | Yes | No | No | Yes | No | None | No |
| J | Cooking school, Thailand | Yes | No | No | No | No | None | No |
| K | Food tours, Cambodia | Yes | No | No | No | No | None | No |
| L | Walking tours, Japan | Yes | Yes | Yes | No | Yes | None | No |
| M | Vespa tours, Vietnam | Yes | Yes | Yes | No | No | None | No |
| N | Food tours, Vietnam | Yes | No | No | No | No | None | No |
| Total (of 14) | 13 | 5 | 3 | 4 | 5 | 0 | 0 |
*robots.txt checked for OAI-SearchBot, ChatGPT-User and PerplexityBot. One operator (D) blocks GPTBot, OpenAI's training crawler, which does not affect ChatGPT search or browsing.
Method. For each operator, one live tour or product page was found by following links from the homepage or sitemap. Each page was fetched as raw HTML, without running JavaScript, as many AI crawlers and simple agents read pages, on 28 October 2026 (about 15:30 UK time). Definitions were strict:
- Price: a price for that tour appears in the page text.
- Real dates or availability: a date picker or availability calendar, specific dates or named departure days, or an embedded booking engine such as FareHarbor, Rezdy, Bokun or Peek. Words such as "departures" or "calendar" on their own, site-wide search filters, review dates and dates on a separate page did not count.
- Booking step in raw HTML: a booking engine embed or link, a checkout or add-to-cart, or a booking form with fields. Contact forms, enquiry links and a bare link to a separate booking page did not count.
- Booking needs JavaScript: no booking step in raw HTML, but the page loads one through a script, a JavaScript modal or a JavaScript-rendered app.
- Structured data: Product, Offer, AggregateOffer, TouristTrip or Event in JSON-LD on the tour page.
Every automated result was then reviewed by hand against the page source, and corrections are recorded with the raw data. Limits: one page per operator on one day; a site can still be bookable by a person, by phone or through a separate booking page; a modern browser-based agent that runs JavaScript may see more. One operator's page (C) returned different content on two fetches, so its result is the later fetch. In total, 24 sites were tried. Besides the 14 included, one returned only a default web server page and seven could not be reached or timed out, so those eight were excluded. Two large activity marketplaces returned HTTP 403 or a bot challenge to the automated request. Bot protection can stop legitimate agents as well as scrapers.
What the results mean
- Prices are mostly visible. Booking is mostly not. 13 of 14 tour pages showed a price, but only 3 had a booking step an agent reading raw HTML can see and use.
- Availability is the weak point. Only 5 of 14 showed real dates or availability on the tour page itself. Without that, an agent has to guess, which is how a traveller ends up with two unplanned days in Ottawa.
- JavaScript hides the booking step. On 4 pages, booking only appears after scripts run, through a widget or modal. Agents that do not run JavaScript cannot see it.
- Structured data is still uncommon. 5 of 14 used Product, Offer or TouristTrip markup.
- Access is not the problem. No operator blocked the AI search crawlers checked, and none had an llms.txt file.
Checklist: make a tour website agent-ready
- Show the price in the page text for each tour, with currency and what it covers (per person, group, private).
- State schedule and availability plainly: days it runs, start times, duration, meeting point, seasonal gaps.
- Keep one clear booking path per tour, with a booking engine link that loads without login and a form with labelled fields.
- Add structured data: TouristTrip or Product with Offer (price, currency, availability) on each tour page, plus Organization or LocalBusiness with contact details.
- Publish operator details: legal name, address, email, phone or WhatsApp/LINE, licence number where relevant.
- Check robots.txt does not block OAI-SearchBot, ChatGPT-User or PerplexityBot, and that bot protection does not reject normal requests to tour pages.
- Keep it current: update prices and dates, and show a "last updated" date. Stale pages are what agents and travellers distrust.
- Add an llms.txt summarising tours, prices and booking links once the above is done.
- Ask how customers found you, and include AI as an option, as Lost Plate did.
NJoy's free AI-readiness check runs several of these tests on any travel website.
Sources
- Travel Massive, How the World's Most Traveled People Use Tech to Travel: Takeaways from Our Panel at Extraordinary Travel Festival, October 2026.
- OpenAI, Overview of OpenAI crawlers.
- Schema.org, TouristTrip and Offer.
- llms.txt proposal.
- NJoy check of 14 operator tour pages, 28 October 2026.