Why AI travel suggestions face a trust test
A holiday recommendation can sound confident even when the traveller has no way to check how it was produced. A conversational assistant may propose a city, rank hotels or build an itinerary in seconds.
That convenience raises a practical question for tourism businesses: what makes someone willing to rely on an AI-generated destination suggestion?
New research suggests that perceived usefulness and trust travel together, but the evidence is more complicated than a simple claim that transparency automatically creates confidence.
A survey rather than a booking experiment
Marko Šostar, Orhan Uludag and Athira Rema Devi Jayakumar examined this question in a study published in Frontiers in Sustainable Tourism on 4 November 2026.
Their evidence came from 227 completed online questionnaires collected in Croatia during May and June 2026.
Respondents were recruited through social media and personal networks, using convenience sampling rather than a representative population survey.
The participants were predominantly young and female.
The study therefore describes relationships among attitudes in this particular group, not the behaviour of every traveller or every tourism market.
What participants were asked
The questionnaire covered five connected ideas. AI and digital familiarity captured respondents’ experience with artificial intelligence and confidence using digital technologies.
Perceived usefulness reflected whether they believed AI could make destination selection more effective. A transparency measure asked how important respondents considered explanations about recommendation processes and personal data use.
Trust assessed confidence in AI information and destination advice.
Finally, behavioural intention asked whether respondents expected to use AI recommendations when planning future travel.
These were self-reported assessments, not independently observed reservations, purchases or completed trips.
How the relationships were analysed
The researchers used partial least squares structural equation modelling, commonly shortened to PLS-SEM.
This statistical method estimates relationships among several measured concepts simultaneously.
It can help assess whether a proposed pattern fits survey responses, including indirect statistical pathways through another construct.
It does not turn a one-time survey into a randomized experiment.
Someone who already likes AI may rate its usefulness and trustworthiness highly for reasons the model cannot fully capture.
The direction of the arrows is part of the researchers’ theoretical specification, not proof that one attitude caused another.
Trust and usefulness stood out
The largest direct relationship with intention to use AI-generated destination recommendations was trust, with a standardized path coefficient of 0.605 and p below 0.001.
Perceived usefulness also had a positive direct relationship with intention, although its coefficient was smaller at 0.259.
The model attributed 65.8% of the variation in reported behavioural intention to its predictors.
That figure describes explained variation within this dataset and model; it is not a claim that AI persuades 65.8% of tourists to book a trip.
Usefulness also related to trust
Perceived usefulness was positively associated with trust, with a standardized coefficient of 0.627.
The analysis also found a statistically significant indirect association between usefulness and behavioural intention through trust, with a coefficient of 0.379 and a 95% bootstrap confidence interval from 0.268 to 0.503.
In plain language, people who rated AI recommendations as more useful also tended to trust them more, and that trust was associated with stronger plans to use the technology.
Because all these measures were collected together, the result should not be described as a proven psychological chain of cause and effect.
Familiarity mattered before trust
General familiarity with AI and digital tools was associated with perceived usefulness, with a standardized coefficient of 0.616.
The researchers also reported significant indirect statistical paths linking familiarity to intention through usefulness, and through usefulness followed by trust.
This pattern is compatible with a common practical experience: people who know what a tool can do may find it easier to recognise potential benefits.
Yet familiarity could also be a marker for broader digital confidence, interest in travel technology or prior positive experiences.
The survey cannot isolate these explanations.
The surprising transparency result
The importance participants attached to AI transparency did not show a statistically significant direct relationship with trust in the fitted model.
Its coefficient was 0.136, with p = 0.113, and the corresponding 95% confidence interval crossed zero.
That finding deserves careful wording.
The study did not randomly show travellers transparent and opaque recommendation systems.
It measured how important respondents said transparency was to them.
Consequently, the result does not establish that explaining AI decisions is useless or that companies should conceal how recommendations work.
Why the distinction changes the headline
Transparency can refer to several different practices: identifying AI-generated material, disclosing data collection, explaining why a destination was selected, or giving a user control over personal preferences. A question about the importance of transparency is not the same as observing the effects of any of those practices.
The paper’s finding concerns one survey construct after other correlated variables were included in the model.
It should not be generalized to actual privacy safeguards, legal disclosure obligations or the performance of explainable AI systems.
Those issues require different tests and evidence.
The model explained attitudes better than it predicted new cases
The researchers checked reliability and discriminant validity and performed sensitivity analyses around the measurement of digital familiarity.
They also tested predictive performance using repeated cross-validation.
Although the model showed predictive relevance compared with a simple mean-based benchmark, a linear-model benchmark produced lower prediction errors for most indicators.
The authors therefore presented the model primarily as an explanation of relationships within the sample, not as a particularly strong tool for forecasting an individual traveller’s future behaviour.
This distinction matters for any business tempted to turn a research coefficient into a marketing prediction.
A measurement choice worth noting
One original item concerning tourism-specific experience was excluded from the final familiarity construct after measurement diagnostics.
The researchers reran the model with that item retained.
The substantive structural coefficients were largely unchanged, which supports the stability of the central associations.
However, the alternative measurement specification had weaker discriminant validity against behavioural intention.
Readers should recognise that modelling decisions can shape how apparently simple concepts such as experience and familiarity are defined.
Sensitivity testing is useful precisely because it reveals whether a headline finding depends on a particular choice.
What travel platforms could learn
For a travel agency or destination platform, the results suggest that a useful recommendation experience and user confidence deserve attention together.
An assistant that offers relevant options, makes it easy to compare alternatives and acknowledges uncertainty may be more compelling than one that merely sounds sophisticated.
But these are design hypotheses rather than interventions tested in the Croatian survey. A product team would need controlled tests to determine whether a clearer explanation, more accurate itinerary or different recommendation interface actually changes trust, bookings or customer satisfaction.
What transparency still requires
The nonsignificant transparency coefficient should not be read as permission to ignore accuracy or accountability. AI-generated suggestions can contain outdated opening hours, nonexistent attractions, unrealistic transport connections or biased recommendations.
Those risks are independent of whether a survey respondent says explanations are important. A responsible service should make critical claims checkable, disclose meaningful limitations and offer straightforward correction routes.
The present study does not measure error rates or demonstrate that its respondents could distinguish accurate from inaccurate AI advice.
A wider question about AI in services
The tourism findings connect to broader questions about technology adoption.
Research Today has previously reported on how hotel guests responded to social service robots, illustrating that consumers may react to digital services in ways that go beyond convenience.
Another report on a redesigned institutional chatbot highlights why actual service experience matters when assessing conversational systems.
These are different studies with different participants and outcomes, so their results should not be treated as interchangeable evidence about tourism decisions.
What the sample cannot establish
The survey was administered in Croatian and distributed through non-probability channels.
It included people with prior travel experience, but was not designed to represent all Croatian travellers, older consumers or international visitors.
The predominantly young, female sample further limits straightforward generalisation.
Its cross-sectional design also leaves room for reverse causation and unmeasured influences. A person intending to use AI could subsequently rate the tool as more useful, while an enthusiastic technology user could give positive answers across several questions.
No randomized exposure to different recommendation systems was conducted.
How to test the idea more directly
A stronger follow-up could randomly assign travellers to recommendation interfaces that vary in accuracy, explanation quality and disclosure.
Researchers could then measure not just stated willingness but verified destination choices, correction of misinformation, time spent comparing options and eventual bookings.
Recruiting participants across age groups, countries and levels of digital familiarity would clarify whether the Croatian patterns recur elsewhere.
Longitudinal data could also reveal whether trust grows after reliable experiences or declines after incorrect suggestions.
These are proposed next steps, not outcomes established by the published paper.
The practical conclusion
The study offers a useful snapshot of how 227 respondents evaluated AI-generated destination advice.
Within the researchers’ model, trust and perceived usefulness were strongly associated with intentions to use such recommendations, while stated importance of transparency did not add a statistically significant direct relationship with trust.
Those findings can inform questions for travel technology designers, but they do not demonstrate that AI recommendations are accurate, that transparency is unnecessary or that trust automatically leads to purchases.
The central lesson is to separate what people say they value, what a statistical model associates with intention and what travellers actually do.
Source Information
Study title: Tourists’ trust and behavioral intention toward AI-generated destination recommendations
Authors: Marko Šostar, Orhan Uludag, Athira Rema Devi Jayakumar
Journal: Frontiers in Sustainable Tourism
Publication date: 4 November 2026
Study design: Cross-sectional online survey analysed using partial least squares structural equation modelling
Sample or scope: 227 respondents recruited in Croatia through convenience sampling, May–June 2026
Key finding: Trust and perceived usefulness were positively associated with intention to use AI-generated destination recommendations; the model explained 65.8% of variance in intention, while the importance attached to transparency had no statistically significant direct association with trust
