There’s been a slow – but increasing – rollout of sophisticated “AI Agents” including ChatGPT’s AI Operator, Manus (coming over from China), and Amazon Nova Act (for developers), which are there to not just do research and make recommendations, but to actually navigate the web and complete actions for users.
NB: This is an article from Triptease
Subscribe to our weekly newsletter and stay up to date
On one hand, AI agents could act as a new way for guests to research travel and find direct websites, acting as challengers to huge OTAs like Booking and Expedia. On the other hand, the OTAs have tech-forward, constantly-evolving booking platforms that may be better equipped to handle traffic from the robotic bookers of the future. Can hotel booking engines keep up?
With this question in mind, we decided to test ChatGPT Operator on 10 of the most common hotel booking engines to see how an AI agent could interact with them – and whether it was able to successfully book a room. Operator is still officially classified as a research project, even though it’s arguably the most advanced option available, so there are still some clear limitations. That fact, combined with some earlier testing, meant we didn’t start with high hopes.
In fact we were somewhat surprised – and impressed – by what we found…
Why we’re keeping an eye on AI Agents
At the time of writing Operator costs $200 per month, so it seems unlikely too many people will be trying it at all yet, let alone paying to do basic tasks like booking a hotel room (some other AI Agents are already cheaper, but are less easy to access). But these tools are going to continue to evolve, and it’s important that direct booking flows evolve with them.
So we’re going to continue to test AI agents to track performance and identify trends or changes. Make sure to check the Triptease blog for your regular updates!
To start us off, let’s see how Operator performed in April 2025..
How Operator performed when asked to book a specific hotel
We set out with a question: are certain booking engines better prepared to handle tasks from AI agents than others? We tested the following booking engines with this question in mind:
- Synxis
- TravelClick
- Selfbook
- Siteminder
- Simple Booking
- SHR
- D Edge
- Travelanium
- Guestline
- Mews
- Webhotelier
For simple booking queries
We used the same prompt for each booking engine:
“Please book me a room at [hotel name] from July 15th to 17th.“
Our initial findings? When given this direct, specific prompt, even without a requirement to book direct, Operator started where most humans would: a search engine. Unlike most humans though, it consistently started its journey on Microsoft Bing, despite Bing’s tiny market share. (Perhaps this has something to do with the amount Microsoft has invested in OpenAI, the creator of ChatGPT…?) Based on all of our testing, if you want Operator to use any other search engine you need to explicitly prompt it to do so.
The rest of the process was good news for direct bookings. After an initial Bing search for a hotel, Operator consistently navigated to the hotel’s direct website. Once there the general flow we saw in each case (with no additional upfront instructions from the user) was as follows:
- Searching for availability at [hotel name]
- Accessing official site, checking availability
- Accepting cookies, checking booking availability
- Accepting cookie policy, clicking book now
- Closing pop-up to continue booking
- Clicking ‘BOOK NOW’ for availability
- Entering dates, checking availability now
- Adjusting dates, setting guest count
- Exploring available room options listed
- Presenting room choices to user
- Selecting [room choice]
- Confirming booking for [room choice]
- Reviewing booking details, proceeding forward
- Confirming booking terms with user
