Travel search is becoming increasingly conversational, and hotels are now trying to secure a place in AI-generated answers. But AI discoverability raises a broader question: which properties become visible, according to what criteria, and through which new forms of intermediation?
“I’m going to Florence for the weekend. I’m looking for an independent, affordable hotel in the city centre, close to the museums.”
Getting an initial shortlist no longer requires opening multiple websites, comparing sponsored results and applying filter after filter. A single request to an AI assistant can return several options within seconds, together with an explanation of why they might fit.
This does not mean that traditional search has been replaced. Rather, another interface is emerging alongside keyword-based search, ranked links and listings: a conversational interface in which users describe what they need and receive an already synthesised response.
The data is beginning to make this shift visible. According to Adyen’s Hospitality & Travel Report 2025, 29% of Italian travellers surveyed said they had used AI to discover destinations, a 77% increase on the previous year. More broadly, Bain & Company reported in 2025 that around 60% of traditional searches ended without users clicking through to another website, while 68% of LLM users were using these tools to search for, gather and synthesise information.
What is changing, then, is not only where people search, but also how information is selected before it reaches them.
In travel, this shift creates a new visibility problem. HotelWorld AI’s AI Visibility Index for the first quarter of 2026 found that only 16% of the hotels analysed appeared in recommendations generated by the AI systems it monitored.
This does not mean that the remaining 84% are universally “invisible” to every model and every query. Visibility depends on the system being used, the way the prompt is formulated, the sources retrieved and the surrounding context. What the figure does indicate, however, is a significant concentration of visibility and the emergence of a new competitive arena.
TAKEAWAYS
Making hotels readable to AI assistants
This is the problem addressed by dAirect, an Italian startup founded by Antonio Picozzi and selected for NVIDIA’s Inception Program. The project has grown rapidly: the company’s website now reports more than 200 hospitality properties served, compared with roughly one hundred during its earlier stage of development.
Picozzi’s starting point is the widening gap between the way many hospitality businesses present information online and the way AI assistants retrieve, interpret and use that information.
“AI has created a layer that sits on top of everything. People no longer move only from one website to another; they can go directly to an assistant. And not just to decide where to stay, but to organise the entire trip: where to eat, what to do, which experiences to choose,” he explains.
The difference from traditional search lies primarily in the structure of the query. A keyword can now become a combination of conditions: dates, number of guests, neighbourhood, price range, family requirements, facilities and the type of experience someone is looking for.
That does not automatically make conventional websites “unreadable” to AI. The issue is more specific. Information spread across multiple pages, poorly structured content, prices and availability stored in separate systems, inconsistencies across sources or data that is difficult to retrieve can all reduce the likelihood that an AI system will interpret that information correctly and use it in a reliable response.
dAirect attempts to measure that likelihood. For each property, hundreds of queries representing different search intents are run in order to observe how frequently the hotel appears in AI-generated answers. That measurement is then used as the basis for technical and semantic optimisation.
There is, however, no stable equivalent of the established practices that developed around SEO. As Picozzi acknowledges, generative models are probabilistic systems: the same query can produce different outputs, while algorithms, retrieval mechanisms and source selection can change over time.
dAirect’s proposed trajectory also extends beyond discoverability. The first stage is to make a property sufficiently interpretable for it to be mentioned in an AI-generated answer. The second is to enable assistants to retrieve up-to-date information such as prices and availability. The third is transactional: allowing an AI agent to complete the booking directly.
This is where visibility becomes a question of intermediation.
Disintermediating OTAs – or creating a new intermediary?
One of the promises associated with this shift is that it could strengthen direct booking and reduce hotels’ dependence on Online Travel Agencies.
The reality is less straightforward.
OTAs begin from a position of strength precisely because they control large volumes of structured content: standardised property pages, pricing, availability, reviews, images and continuously updated information across very large numbers of hotels. These features are valuable not only to traditional search engines but also to systems that need to retrieve information in order to construct generated answers.
AI assistants could therefore reduce certain forms of intermediation without necessarily reducing intermediation as a whole. The actors may change, as may the points of access and the mechanisms through which visibility is distributed.
dAirect itself occupies an interesting position in this transition. It builds a technical layer between hospitality providers and AI ecosystems with the aim of helping hotels reduce their dependence on traditional intermediaries. The apparent contradiction is revealing rather than problematic. Digital disintermediation often does not eliminate intermediaries; it changes their function.
The popularity-bias problem
There is also a second issue, less visible but potentially more consequential: how are the hotels that appear in recommendations actually selected?
Luca Pappalardo, a researcher at ISTI-CNR in Pisa, has long studied popularity bias: the tendency of recommendation systems to favour items that already enjoy higher visibility or greater levels of interaction.
The mechanism can generate a feedback loop. More established properties are recommended more frequently; those recommendations generate further interactions; and those interactions reinforce their original position.
“Even when this effect is unintended and the aim is simply to use the information available, similarity with other users or with places that have already been visited can still produce a ‘rich get richer’ dynamic,” Pappalardo explains.
With generative assistants, the problem takes on different characteristics from those found in traditional recommender systems. Not all LLMs operate in the same way, and some may incorporate location, browsing or interaction history, memory, real-time information or external search systems. Where the available context is limited, however, online prominence, the volume of accessible information and representation within retrieved sources may carry considerable weight.
The risk is that a system that appears capable of expanding choice could instead narrow it around the same actors that are already strongly represented across the digital environment.
That outcome is not inevitable. It is a dynamic that needs to be measured.
Pappalardo himself points out that correcting popularity bias is not simply a matter of recommending less well-known options. Doing so might reduce the relevance of recommendations for users. The challenge is therefore to balance accuracy, diversity, distribution of visibility and quality of experience. Research into how these competing objectives can be reconciled remains ongoing.
From SEO to AI discoverability
The nature of uncertainty is also changing for businesses.
The rules governing search-engine rankings have never been fully transparent, but SEO has gradually developed a body of shared knowledge: technical documentation, metrics, professional communities, analytical tools and accumulated experimental practices.
AI discoverability is far less stable.
An LLM may rely on knowledge embedded in the model itself, search engines, databases, APIs, retrieval systems and other external sources. Different systems may also construct their answers in very different ways.
The term RAG – Retrieval-Augmented Generation – also requires some clarification. It does not simply describe “the technology through which an LLM cites its sources”. It refers to an architecture in which the model retrieves external information before generating its response. Citations may be part of the user interface, but they are not what defines RAG.
Pappalardo also points to the possibility that some large platforms may be disproportionately represented among the sources retrieved by AI systems. Actual source distribution, however, varies according to the model, the search infrastructure involved, the query and the architecture of the service. There is currently no universal hierarchy of sources that applies consistently across all assistants.
For a hotel, this creates a new kind of problem. The question is no longer simply how much to invest in visibility, but which forms of visibility are worth investing in and how their effects can be measured.
“Because the system is a black box, there is still no clear answer,” says Pappalardo.
Who can afford to be visible?
The issue therefore also has a distributive dimension.
A large hotel group will typically have structured websites, sophisticated booking systems, in-house digital expertise and the budget to experiment with new optimisation tools. Independent hotels, agritourism businesses and family-run properties may operate with very different resources.
If visibility within AI environments increasingly requires new levels of data structuring, monitoring, experimentation and technological integration, the gap between operators could widen.
Not because AI assistants are necessarily designed to favour larger organisations, but because digital readability is itself the product of infrastructure, expertise and investment.
That distinction matters. A system can be formally neutral with regard to company size while still producing non-neutral distributive effects if some organisations are substantially better equipped than others to satisfy its requirements.
When visibility can be bought
Another variable now needs to be added to this asymmetry: the commercial model of AI assistants.
During 2026, OpenAI expanded its advertising programme in ChatGPT to several markets, while Google already allows advertisements to appear above, below and, in some countries, within AI Overviews.
This does not mean that organic recommendations produced by AI assistants will necessarily become a new market for keywords. OpenAI, for example, explicitly states that advertising is kept separate from generated answers.
It does mean, however, that monetisation is moving into conversational interfaces as well.
For the travel industry, the distinction will become increasingly important: how much visibility will be determined by recommendation systems, how much by the quality and accessibility of data, and how much will ultimately be available for purchase?
The history of online search suggests that this balance should not be treated as settled.
More context, less exploration?
The perspective of the traveller also remains largely outside this debate.
AI assistants allow people to formulate requests that are considerably richer than a conventional keyword and to narrow a large field of alternatives very quickly. That is precisely what makes them useful.
But the same compression of choice can also have side effects.
A conventional search-results page exposes users to multiple possibilities at once. A conversational answer, by contrast, tends to produce a shortlist and may occasionally converge on a single recommendation. If the user accepts that selection without exploring further, there is less opportunity to encounter alternatives that had not previously been considered.
This does not mean that serendipity disappears. An assistant can be explicitly asked to identify unusual, less popular or unconventional options. What changes is where the act of selection takes place. Part of the exploration previously undertaken directly by the user is delegated to the system.
That makes the criteria behind the selection more important.
More regulation, but still limited evidence on the effects
Compared with the early stages of this transition, the European regulatory framework has meanwhile become more operational.
The European portal established under the Digital Services Act to allow qualified researchers to request access to data held by Very Large Online Platforms and Very Large Online Search Engines was launched in July 2025, with the associated data-access regime becoming fully applicable on 29 October of the same year.
A further phase of the AI Act also entered into application on 2 August 2026. A number of obligations became applicable and enforceable, including provisions concerning general-purpose AI models and new transparency requirements for certain AI systems.
These mechanisms may increase the capacity to observe and govern certain aspects of the ecosystem, but they do not automatically answer the central question raised by research into recommendation systems.
“We are still at the level of intuition,” Pappalardo says. “We need rigorous analysis to understand what happens in a feedback loop when a traditional recommender system is replaced by a large language model.”
This is perhaps the most significant aspect of the transition now under way.
Competition for online visibility is not simply moving from Google to ChatGPT, Gemini, Claude or other systems. The architecture through which information, recommendations and transactions are organised is itself changing.
Startups such as dAirect are developing tools designed to help hotels operate within this emerging infrastructure. At the same time, researchers and institutions are still trying to understand its distributive effects, its feedback loops and the forms of market concentration it may create.
The central question, then, is not simply how to make a hotel visible to an AI assistant.
It is how the criteria governing that visibility are being defined, who is best positioned to adapt to them, and what forms of intermediation may emerge as search becomes increasingly conversational.
