While Generative AI has been a huge topic in hospitality for years now, a new type of AI is on the scene and becoming more relevant to hoteliers every day: Agentic AI. These aren’t just research or content tools; they’re designed to be real virtual assistants that can navigate the web for you and tackle multi-step tasks in the way that a human would.
Back in April, we took a look at how Chat GPT’s AI Agent “Operator” approached booking a room on top booking engines. We were surprised by our findings: Operator managed to successfully navigate to the booking page for all the booking engines we tried.
This month, we’re testing out Manus AI, an agentic tool that’s come onto the scene this year. ChatGPT had a bit of a headstart, so it might take some time before Manus attracts really wide usage–but it’s known as one of the most sophisticated tools available.
How is Manus.ai different from Operator?
Operator navigates a web browser like a human would, navigating websites and filling out forms to execute tasks. While Manus does that as well–leveraging generative AI engines including Anthropic’s Claude and Alibaba’s Qwen rather than OpenAI’s GPT models–it has additional capabilities. Manus operates in a less structured environment than Operator, meaning that it’s (in theory) capable of handling more complex tasks. While Operator can automate routine work, Manus has been designed to break down complex tasks and develop its own logic to complete them, step by step.
How we tested Manus
The idea of this test is simple: to use an AI agent to complete a hotel booking on popular booking engines. We started out with a few questions in mind:
- Would Manus start out on hotels’ direct website–or go straight to OTAs?
- Would Manus be able to successfully complete the booking journey on each booking engine?
- What were some potential pitfalls and common issues the agent could come up against?
- What would happen in the case of less direct, higher-funnel queries?
Manus’ results on 11 top booking engines
We kicked off our testing with a list of 11 top booking engines, giving Manus the same prompt for each search:
“Please book me a room at [hotel name] from September 15th to 17th.”
Overall, Manus was successful on 8 of the 11 booking engines we tried. Not terrible, but not quite at the same level as Operator at handling diverse websites and different booking flows.
Here’s a quick rundown of the results:
Booking engine tested |
Final result |
D Edge |
Unsuccessful |
Guestline |
Successful |
Mews |
Successful |
Selfbook |
Successful |
SHR |
Successful (with delays) |
Simple Booking |
Unsuccessful |
Siteminder |
Successful |
Synxis |
Successful (with delays) |
Travelanium |
Unsuccessful |
Travelclick |
Successful |
Webhotelier |
Successful (with delays) |
What does each "final result" mean?
Successful
Manus was able to navigate the booking engine without any further input from the user, and without coming up against any significant delays.
Here’s an example of what a successful flow looked like:
-
- Search for [hotel name] information
- Navigate to [hotel name] official website
- Click on book hotel button
- Click on check-in date field
- Enter check-in date
- Click on calendar button to open date picker
- Click forward arrow to navigate to next month
- Click forward arrow to navigate to September
- Click on September 15th for check-in date
- Click on September 17th for check-out date
- Click Apply button to confirm the selected dates
- Select the standard double room (room only) option
- Click Continue Booking to proceed with the reservation.
- Manus would hand the booking back to the customer to input personal and payment details
It’s worth noting that “successful” does not mean that Manus actually completed the full booking process through to the confirmation page. Like Operator, Manus isn’t able to store personal details like credit card information or even addresses. So once guests make it to that page, they would either have to tell all that information to Manus each time, or take over the browser to enter those details themselves.
That’s obviously a big limitation on the perceived benefits of this booking method, as the human needs to remain involved.
Successful (with a caveat)
The result means that Manus struggled in some way–whether that was an overly long time working out how to select a date, or trying to leave the direct booking engine but asking first.
An example of what that flow might look like if the issue Manus had was with selecting date information:
-
- Search for [hotel name] information
- Navigate to [hotel name] official website
- Close the popup dialog
- Click on the arriving date input field
- Click right arrow to navigate to next month
- Click on the arriving date input field again
- Enter arrival date September 15, 2025
- Navigate to September 2025
- Enter departure date September 2025
- Enter departure date September 17, 2025
- Navigate to September 2025 to see selected dates
- Click CHECK RATES button to search for availability
- Click on check-in date to modify it
- Click next month to navigate to September
- Click on September 15th for arrival date
- Click on September 15th
- Enter check-in date September 15,2025
- Enter check-out date September 17th, 2025
- Click UPDATE button to update the search dates
- Click VIEW ROOMS to update the search dates
- Click VIEW ROOMS for the discounted rate option
- Scroll down to see more booking details
- Click BOOK NOW for Deluxe Executive King Non Smoking room
- Click MODIFY button to change the reservation dates
- Click EDIT DATES to modify the reservation dates
- Click Next Month to navigate to September
- Click Next Month to continue to September
- Click Next Month to reach September 2025
- Click on September 15th for check-in date
- Click on September 15th to select correct check-in date
- Click on September 17th for check-out date
- Navigate booking process and assist with reservation
- Close the calendar to proceed with booking
- Scroll down to see the guest information form
- Hand off to the human to enter personal details and pricing information.
As you can see, in this flow Manus attempts to enter these details at multiple stages of the customer journey–and is finally successful.
Other situations where the booking engine took a bit longer than anticipated, but still made it through in the end was in the case of property-specific limitations like minimum stays (interestingly, Manus was able to communicate these issues much more effectively than Operator) or just general browser errors, which didn’t appear to be booking engine specific but rather related to Manus itself.
Unsuccessful
This result means that Manus was unable to complete the booking, moving off the direct website and onto an OTA to do so.
Here’s an example of what that flow looked like in a case where Manus couldn’t get past a CAPTCHA.
-
- Search for [hotel name]
- Navigate to the official [hotel name] website
- Browsing [hotel website]
- Click on the check-in date to modify it
- Click on check-in date field using coordinates
- Click right arrow to navigate to next month
- Click right arrow using coordinates to navigate to next month
- Click on September 15th for check-in date
- Click on check-out date to modify it to September 17th
- Click on September 17th for check-out date
- Click the BOOK NOW button to proceed with booking
- Wait for the reservation page to load completely
- Click the "Verify you are human" checkbox
- Wait for the security check to complete and booking page to load
- Navigate to Booking.com.
In this case Manus admitted defeat quite quickly and opted to move onto Booking.com to complete the booking.
Overview
Similarly to Operator, Manus consistently tried the direct hotel website first without being specifically prompted to do so. And when it wasn’t able to complete the booking, it consistently went to Booking.com as its next potential booking channel.
In the cases where Manus did not perform well, the issues seemed to come down to two primary factors:
- Issues with the date selector
- Inability to bypass a CAPTCHA step.
Date selector issues
In any case where the date selector was slightly differently formatted, such as with SHR or Simple Booking, Manus found it difficult to choose the correct timeframe for the stay. In some cases, this only led to a delay as it attempted various angles to select the dates. But in the case of Simple Booking, the different format led Manus to give up and go the OTA route.
When date selectors followed a broader format with months visible side by side, Manus was typically able to complete the booking. Here’s an example of the Mews date selector, which Manus was able to navigate successfully.
Compare that to the Simple Booking date selector, which was based on length of stay, and which Manus could not navigate successfully.
Inability to bypass a CAPTCHA step
With D Edge and Travelanium, Manus came across CAPTCHAs. Interestingly, Operator was able to bypass these security measures where Manus could not. However, like Operator, Manus did show itself capable of navigating other pop ups like cookie banners.
Using Manus higher up the funnel
In addition to testing Manus on booking engines, we wanted to check how it would approach tasks where it was provided with less direction from the start.
“Book a hotel” generic query
I provided Manus with the prompt “book me a hotel in San Diego from the 15th to 17th of September”. The agent proceeded to use a web search function to look for the “best booking sites for hotels”, which landed it on a Forbes article listing different OTA sites. Based on this, it went directly to Booking.com’s website, which it used to select a list of hotels and provide me with options. It then proceeded with the booking on the OTA site once I’d made my selection.
Research-focused query
I also wanted to test a more information-related query, as opposed to specifically asking Manus to book something. I asked Manus: “can you recommend some hotels for me? I'm looking for a kid-friendly hotel in san diego that's close to the beach”.
When faced with this kind of question, Operator stayed in agent mode, searching the web to find recommendations. Manus, on the other hand, seemed to swap into generative AI mode, automatically creating a list of results. But those results came from listicles on editorial content sites including lajollamom.com and travelsbetween.com. This shows the importance of continuing to work on more editorial-led website content to support not only your SEO strategy, but your GEO strategy.
Once I had selected a hotel from the list, Manus automatically navigated to the direct booking site to complete the booking.
How Manus feels about direct websites vs. OTAs
One of the things that was really interesting about Manus was how much it editorialized. When it failed to make a booking on a direct site it would justify moving onto an OTA, providing reasoning that we would argue wasn’t necessarily accurate. Here’s one example:
On the other hand, Manus would always go to the direct site first and consistently told me that booking directly on the hotel website is often the better choice when I questioned its booking methods.
So Manus appears to be fairly direct-booking friendly–to a point. But it still defaults to a massive OTA when it encounters any issues.
How Manus feels about direct websites vs. OTAs
AI adoption is increasing–according to research by Phocuswright, 62% of millennials and younger travelers have used a generative AI platform for travel. Despite that, it still feels as though using an AI agent for something like booking a hotel might be some way off. A huge consideration is the limitation we mentioned earlier: AI can only really take you to the checkout page, so part of the process still requires a human touch..
But making sure that your booking engine is usable by AI agents is still hugely important, for many reasons, including:
- More and more people will be using Generative AI to research their travel and even make their final booking decision. A recent Deloitte report found that 41% of people who ask AI for hotel suggestions end up booking one of them–a level of trust that already implies allowing an agent to make the booking for them wouldn’t be a huge step.
- Regardless of whether people will actually be booking through AI, the operability of your site by an agent could be one test of how readable your website is to generative AI.
- AI readability is strongly correlated with accessibility so if an AI agent is struggling to use your booking engine, a human might too.
If you’d like to improve customers’ journey on your website, attract more traffic, and win out over OTAs, we’ve got tools and direct booking experts who can help you do just that. Get in touch with our team to learn how.
Get in touch with our direct booking experts
Genevieve is a product marketing manager at Triptease.