Beyond Museums and Monuments: The Travel Preferences That AI Guides Can Easily Miss

The study examines whether small pictures, photographs and an AI reading of visitors’ words reflect how people describe their own travel interests. Its findings raise questions about recommendations that appear personal because they resemble what many travellers already enjoy.

A perfect day in a historic city might include ancient ruins, a neighbourhood café and hours of wandering without a timetable. A travel app could recognise your interest in history and food, then fill every hour with stops that leave little room for discovery. Understanding the places you enjoy captures only part of what makes a trip feel right.

Researcher Markos Konstantakis explores this gap in "What Visitors Say and What Icons Show: Re-Examining a Cultural-Visitor Typology for Generative-AI Personalisation in Heritage Guides," published in Heritage. The study examines whether small pictures, photographs and an AI reading of visitors' words reflect how people describe their own travel interests. Its findings raise questions about recommendations that appear personal because they resemble what many travellers already enjoy.

The exploratory survey involved 37 adults, including 33 living in Greece, who interpreted six icons, selected interests and photographs, and described an ideal day. Participants rated themselves against eight visitor profiles, giving the researcher a reference for comparing their answers with profiles inferred through different inputs. A Gemini AI profiler analysed their written descriptions using fixed instructions, allowing its results to be checked across repeated runs.

Small Pictures Can Send the Wrong Message

An icon may look straightforward to its designer and suggest something quite different to the person tapping it. Only the nightclub and distillery symbols conveyed their intended meanings to more than two-thirds of participants, with recognition reaching 81% and 70%. Recognition fell to 54% for the brewery, 43% for places of worship, 16% for the cultural centre and 5% for the handmade musical instrument workshop.

The misunderstandings followed recognisable patterns rather than random guesses, revealing how familiar associations shape what people see. A Star of David intended to represent all places of worship suggested a synagogue or Jewish heritage to 16 respondents, matching the number who chose the intended meaning. The instrument workshop suggested a museum or traditional music, and the cultural centre often looked like a concert venue or opera house.

Some mistaken readings still pointed towards the intended broad interest, allowing a guide to recognise an enthusiasm for music without identifying the right kind of place. Such confusion can change the experience someone receives and narrow how religious or cultural heritage is represented. The paper recommends adding clear labels and testing symbols with intended users, since smartphone familiarity offered no detectable advantage in interpreting them.

Travellers Have Mixed Interests and Their Own Visiting Style

The study used the ACUX framework, which includes Archaeologist, Art Seeker, Leisure Seeker, Religious Seeker, Naturalist, Traditionalist, Gourmand and Viral Seeker. Thirty-five participants strongly identified with at least two profiles, and the typical respondent rated four descriptions highly. Attempts to sort people into distinct groups found no clear structure, supporting profiles that preserve a person's mixture of interests.

History and famous attractions attracted strong ratings, with religious travel receiving the lowest average rating in this sample. The Viral Seeker category, described through famous highlights and currently popular attractions, attracted high self-ratings that its pictures and interest tiles did not closely reflect. Its wording may describe ordinary sightseeing too broadly to identify a distinct preference, making a clearer definition necessary.

Visitors' written accounts exposed something the eight categories barely captured: how they wanted their day to unfold. Twenty respondents described walking, spontaneous exploration or less touristy places, and 12 mentioned everyday local life. Among answers about things to avoid, crowds, queues and peak season appeared most often, followed by touristy places and rigid planning.

Konstantakis proposes testing an additional layer covering pace, planning, local immersion, walking and tolerance of crowds. These preferences could shape the length and rhythm of an itinerary as well as the places it includes. The proposed layer remains a hypothesis because one researcher coded the themes, and asking about an ideal day naturally encourages descriptions of visiting style.

AI Recognised the Average Visitor Better Than Individual Differences

A generic profile based on other participants' average ratings matched visitors' self-descriptions at least as well as any of the three profiling methods. This reveals a problem for judging personalisation, because a system can appear convincing by repeating preferences shared across a group. The researcher removed that shared pattern to examine how much each input captured about individual differences.

Interest pictograms ranked first in this comparison, photographs came next and AI interpretations of free text ranked last. The differences did not remain convincing after adjustments for multiple comparisons, so the study cannot establish that one method is better. Each approach retained some information about individuals, and none reliably placed the visitor's chosen best-fit profile first.

The AI reproduced the group's average ordering of interests most closely and selected the same top profile in 33 of 37 cases across two runs. That consistency showed repeatability, without demonstrating that each person had been understood correctly. Longer descriptions were associated with better individual agreement, an important detail given that the typical response contained only 32 words.

The comparison involved different kinds and amounts of information, which limits how directly the results can be interpreted. Participants chose from 40 interest tiles and 16 photographs, and the writing question invited a description of activities rather than a direct statement of preferences. Photographs could attract attention for their beauty, and the AI tended to give local food and traditional life more weight than the visitors' ratings suggested.

Better Guides Need Explanations and Room for Correction

Participants were receptive to digital assistance, with 31 expecting an app to be very helpful and 28 expressing those levels of trust in AI suggestions. Explanations mattered to 36 respondents, showing that enthusiasm for recommendations came with a strong desire to understand them. Written descriptions were the most commonly selected information people would share, ahead of interest icons, location and photographs.

The paper recommends showing visitors their mix of interests, explaining how it was inferred and allowing corrections. Combining text with explicit choices and asking follow-up questions could help address incomplete descriptions or uncertain interpretations. Responsible systems should record their instructions and outputs, explain how submitted words are processed, and avoid treating an interest in religious sites as evidence of personal faith.

These findings come from a small, highly educated group recruited through professional networks, with 34 participants answering in Greek. Self-ratings were a useful reference rather than an independently verified measure, and the AI findings concern one model and one set of instructions. Larger, more diverse studies are needed to test labelled icons and visiting styles, with this research offering a practical starting point for guides that give travellers more say in how they are understood.