AI still travels with the handbrake on in hospitality

Asking an AI to find a holiday apartment already feels natural. We provide a destination, dates, guest numbers and a few preferences. We expect the assistant to search for properties, compare availability and tell us what we will spend.

Part of this journey already works. Another part can stop at the most important moment: checking whether the accommodation is available and obtaining the price for the requested stay.

AI has learned to find accommodation. It does not yet have full freedom to discover its price.

At Lestis, we have two ways of observing what is happening: the traffic AI systems generate on websites hosted on our infrastructure, and the practical experience of people using them to organise a trip.

The data shows when and how often, relative to other requests, these systems seek pricing information. Tests with ChatGPT, Gemini and Claude show how difficult the transition from reading a page to checking real availability can be.

AI looks for prices, especially in spring

Lestis Intelligence Observatory analyses traffic on websites hosted on our infrastructure, distinguishing human users from automated systems and classifying AI agent requests by intent.

One category tracks Prices and Availability. Others cover content extraction, indexing and learning, and resource existence checks.

What does “resource existence check” actually mean? Before exploring a website, automated systems look for technical guidance to help them find their way: which pages are available and what instructions the operator provides for automated access. It is rather like looking for a map and house rules at a property's entrance before exploring its spaces. These requests concern how the website is organised and can be explored; on their own, they do not indicate that an AI is reading an apartment description or looking for a stay's price.

This preliminary activity accounts for a significant share of observed traffic. To understand the share devoted to prices and availability, it helps to examine these requests both as a proportion of all AI traffic and with technical checks excluded.

The third column is recalculated from the Observatory's published percentages, dividing the pricing share by the share of requests remaining after resource checks are excluded. Results are rounded to one decimal place.

Across all AI traffic, the peak is in April. Excluding resource checks shifts it to May.

In May, almost one in five AI requests other than resource checks concerned prices and availability. In October, the share was 5%.

The distinction also matters after summer. Between August and September, the pricing share of total traffic halves, from 3.2% to 1.6%. Excluding checks, it falls from 6.5% to 5.9%: a much smaller decline.

Part of the apparent fall therefore reflects the increasing weight of technical checks. The decline from spring remains clear even after they are excluded.

These percentages describe the composition of requests, not their absolute volume. A smaller share does not automatically mean fewer pricing requests.

ChatGPT, Gemini and Claude: they find the accommodation but cannot reach a quote independently

We tested ChatGPT, Gemini and Claude on finding a stay's price on websites built with Lestis. The common limit was this: assistants could find the property and read published information, but could not independently complete the step of checking prices and availability.

The tests used the standard chat interfaces, asking the assistants to search for a stay through a natural-language conversation: the way people commonly use these tools around the world.

The context matters: price discovery and verification tests concerned Lestis websites, where we apply all known industry best practices to make content and information accessible and understandable to automated systems. On less optimised websites, assistants might have even greater difficulty finding accommodation or reading its starting price. Equally, better results elsewhere cannot be ruled out.

With ChatGPT, using a paid subscription, we asked it to search for a stay and check its cost. The assistant found the website, located the page for the unit we wanted to rent and read the base daily price. When it needed to check the dates, however, it could neither submit the availability search form nor read the page's structured data to use it in the query.

Opening a direct link containing search parameters also ran into page retrieval limitations, known as fetching. In the session tested, useful information remained accessible through search engine results and indexed or cached copies, without reaching an up-to-date response from the booking system.

A step forward with Claude! As well as finding the property, its features and daily price, Claude could read the result and quote when we supplied the direct link. When we asked it to construct a new combination, it reported receiving the homepage or a previously read page and could not view the quote, even though the page it was trying to open was correct. Paste the link into the chat and it could view it; otherwise, it could not.

Pasting the link allowed Claude to read the quote. But this scenario does not solve the problem.

We asked Gemini to find the actual price of a stay for two adults from 1 to 7 March, following the same process as the other tests. Gemini also found the property's website and the accommodation's starting price, but did not return a quote for the requested dates. To see the actual cost or check availability, it told us to click the link and continue on the website. The decisive step remained the user's responsibility.

Three assistants, the same obstacle: independently moving from published information to the booking system's response. In all three tests, chat searches found the accommodation and starting price, but stopped before obtaining availability and a quote without user intervention. With Claude, supplying the direct link overcame the block and allowed it to read the result.

We were not asking it to complete a payment or confirm a booking. The information we started searching for was still missing: what that stay actually cost, for those dates and those guests.

The tests concern the environments and tools used: operational browsers, APIs and dedicated integrations can offer different capabilities. The difficulty encountered even with a paid ChatGPT subscription shows, however, that a subscription alone does not guarantee operational access to a property's booking system.

A starting price is not the price of a holiday

An apartment's price depends on arrival and departure dates, length of stay, guest numbers, rate conditions and actual availability. Mandatory charges, additional services and minimum-stay rules can also affect it.

Reading a starting daily rate is therefore not enough to produce a reliable quote. Multiplying it by the number of nights risks giving a result the booking system would not confirm.

Answering the traveller's question requires querying the booking engine, which calculates availability and rates for the requested stay, and reading its current result.

This is where the gap between understanding an offer and verifying it becomes clear.

A website can expose descriptions, structured information and machine-readable booking actions. But the assistant must be able to access that information, construct the request and read the response. If navigation tools prevent any step, the journey stops.

Website readiness and the assistant's operational freedom are two conditions that must come together.

Why does price searching rise until May and then decline?

Another aspect of the Lestis data deserves attention: the possible connection with traditional seasonality in holiday searches. If travellers start involving assistants in choosing accommodation, AI requests might also reflect the calendar they follow when organising their stays.

The peak in pricing requests falls between April and May, when many tourists choose and book their summer holidays.

Property operators recognise this pattern. Enquiries for summer start arriving in February and March. In April and May, many travellers move into more concrete planning: comparing destinations, assessing apartments and checking the cost of possible options.

It is a cycle many hosts know through enquiries and quote requests: first choose where to go, then compare accommodation and prices, and finally book. AI requests show an interesting resemblance to this journey and raise a question: is AI-assisted holiday searching beginning to show seasonality of its own?

Excluding resource checks, the pricing share rises from 2.9% in February to 7.8% in March, reaches 17.3% in April and 18.2% in May. Relative interest in prices and availability grows just as summer holiday choices become more concrete.

One possible interpretation is that some of this planning already takes place through AI assistants. If people consult them while preparing holidays, their requests to websites might also concentrate in the months when travellers compare offers.

AI can search for properties, visit accommodation pages, identify starting prices and try to reach availability, even when it cannot complete a real query to the booking system.

Even with the handbrake on, it tries to do what tourists ask it to do.

After May, the picture changes. In June, the pricing share excluding checks falls to 7.3%. July and August stay relatively close, at 6% and 6.5% respectively.

This pattern also makes sense to hospitality operators. As summer approaches, some main holiday bookings have already been made. Searching continues, including for last-minute stays, but advance planning gradually gives way to departure and the holiday itself.

This could offer a natural explanation for the decline from June onwards and relative stability during the central summer months.

In September and October, the share falls again, to 5.9% and 5%. The decline is more gradual than total-traffic percentages alone suggest, because resource checks gain weight over the same period.

Taken together, the picture recalls traditional summer holiday planning seasonality: growth during preparation, a decline once some stays have been booked, and further adjustment after high season. This parallel could be an early indication of seasonality in AI searches too, linked to the timing and needs of the travellers using them.

Timing alone does not prove cause and effect. The data describes AI system requests, not individual travellers' intentions; seasonality also varies with destinations and the characteristics of the properties monitored.

The signal deserves attention: traffic we tend to think of as purely technological might already be following, at least partly, the rhythm of our holidays.

What changes for a property's website?

For hosts, these developments concern a practical issue: their own website's role in direct distribution.

The familiar journey starts with a search, continues with a website visit and leads to checking prices and availability. With AI assistants, some steps can happen on the traveller's behalf before they even open the property's page.

It therefore matters that rooms, apartments, services and conditions are recognisable, and that a clear route to actual availability exists.

Hospitality has long discussed SEO and, more recently, content optimisation for generative search engines. Accurate descriptions, structured information and accessible content help an offer be understood. Commercial queries add another step: checking whether that offer can be purchased on the requested terms.

Recognising that a property has three apartments and understanding their amenities is one thing. Asking which are available for two people from 10 to 17 August and receiving the stay's rate is another.

At Lestis, the website, price and availability management and direct booking engine are part of one system. This connection is central to taking guests from discovering accommodation to checking a commercial offer, and provides a foundation for developing interactions with AI assistants.

Assistants' ability to use the information still depends on available tools and integrations. That makes observing visits and testing real search journeys as important as preparing content.

Lestis is working to be ready when the handbrake is released

The 2026 data describes a phase in which pricing queries exist, show a pattern compatible with tourism seasonality and coexist with operational restrictions on accessing information.

Tests with ChatGPT, Gemini and Claude show where the journey can break down: the assistant finds the apartment, reads its presentation and base price, but cannot independently verify the requested stay. Gemini requires continuing on the website; with Claude, pasting the right link made the quote readable. In both cases, a decisive step remains with the user.

When checking availability and rates becomes more widespread and reliable, a new opportunity will open for properties' direct channels. Assistants will be able to compare actual offers and guide travellers towards an offer verified on the property's own website.

Lestis works on and monitors these developments to be at the forefront when this opportunity can be taken.

The Observatory, traffic analysis and assistant tests help us understand what already works and where the journey stops. We continue following the evolution of tools and access methods, aiming to put new possibilities to work for our hosts' websites and direct bookings as soon as they become practically usable.

That work starts today: observing how AI searches for accommodation, understanding which information it reaches and preparing the ground for the next step.

Disintermediation is at stake. If AI assistants can query official websites' booking engines directly, including those built with Lestis, they will be able to find and compare current offers at the source and guide travellers towards booking with the property.

For direct bookings, this could be a revolution: an opportunity to reduce dependence on OTAs such as Booking and Airbnb, rebalance their power in distribution and bring value back to the relationship between guests and properties.

AI is already looking for accommodation for our holidays. Lestis is working so that, when it can reach the quote, property websites will be ready to respond.

And you, are you ready?

---

Data source: Lestis Intelligence Observatory, monthly data from February to October 2026. Shares excluding resource checks are recalculated from published percentages and rounded to one decimal place. The seasonal interpretation is an analytical hypothesis. The ChatGPT, Gemini and Claude cases describe direct tests using standard chat interfaces on Lestis websites. ChatGPT was tested with a paid subscription, and the Claude test was checked against server logs.