The Concentrated City: Effects of AI-Generated Travel Advice on the Spatial Distribution of Tourists

Abstract

The analysis of the spatial location of tourists is essential for effective tourism management. This study explores the potential effects of large language models (LLMs) on urban travel planning. Despite growing academic interest in LLMs, empirical research on their specific impact on urban tourist locations remains limited, even though these models may significantly affect tourist behavior and spatial dynamics. This article compares the location of heritage sites in the city of Barcelona that are traditionally visited by tourists (as identified through Instagram) with those recommended by ChatGPT. The results show that ChatGPT tends to recommend a much smaller and more spatially concentrated number of tourist attractions than those shared on Instagram. The findings indicate that ChatGPT reinforces mainstream representations of cities by prioritizing well-known landmarks, potentially overlooking emerging or local attractions. This simplification can lead to tourist overcrowding and the marginalization of less-visited areas. Likewise, it may entail new needs for the management of urban spaces. Urban planners and tourism managers may need to intervene to redistribute tourist flows in a context where various models of tourist behavior will coexist.

1. Introduction

The emergence of large language models (LLMs), powered by artificial intelligence (AI), has brought about a significant shift across various sectors, including tourism. A study conducted in the United States and Canada indicated that more than one third of leisure travelers had already used LLMs within only six months of their release [1]. The same study also observed that 84% of users expressed satisfaction with the results obtained. Furthermore, 54% of leisure travelers currently trust generative AI for providing travel ideas and helping with planning, with 81% indicating their willingness to make bookings based on its recommendations. On this basis “AI will become almost a de facto standard for how [some travelers] think about their travel plans and how they even dream about their vacations” [1]. The academic literature has also noted that “tourist–chatbot interaction would increase destination image and visit intention” [2] (p. 1). However, the information that language models provides may also be biased. It is important to understand the kind of information produced by AI in order to understand the nature of the information a tourist will obtain. This can provide insights into the locations likely to be visited by tourists, with the understanding that such expectations could contribute to the more effective planning of tourist areas.

However, despite the increasing use of AI-powered language models (LLMs) in the tourism sector, their impact on the spatial distribution of tourist flows remains significantly underexplored [3]. While previous studies have focused on the effects of digital media and user-generated content on destination image and perception [4,5], relatively few have addressed how LLMs might shape actual tourist behavior, particularly in terms of the places they choose to visit [6]. This paper seeks to fill this gap by analyzing the spatial recommendations made by LLMs. In this way, it aims to clarify how generative AI may contribute to reshaping tourism geographies in urban contexts.

The image of a tourist destination was traditionally based on a limited number of locations disseminated by institutions or companies [7,8]. Generic images gave the tourist an illusory sense of having understood everything, and of order and stability: a feeling of having everything under control [9]. Gradually, tourists have transitioned from merely consuming the image of a place to actively shaping it themselves; the images they construct through personal perception are frequently regarded as more trustworthy than those circulated by official institutions [4]. This process has been described by several academic studies that have highlighted the differences between the projected image and the perceived image from various perspectives, including the image creation process itself [10], its subject matter [11], and user experiences [12]. More recent scholars have distinguished among three key concepts [13]. First, place image is defined as the subjective perception of an identity that is objectively established. Second, place identity refers to the collection of elements and attributes that characterize and differentiate a specific territory from others. Third, place reputation is conceptualized as a consensual evaluation formed by external audiences, resulting in a collectively accepted dichotomous judgment—typically categorized as either a “good” or “bad” reputation.

The emergence of LLMs poses a significant new challenge. In order to optimize their responses, LLMs must collect data from extensive sources of information, both projected and perceived. This implies providing tourists with a new source of information. The new functionalities offered by LLMs pose a significant challenge when managing tourist flows. A number of studies have shown how tourists who are exposed to too much information on websites and social media may develop a negative perception of a particular destination [14]. The capacity of LLMs to collect and process large amounts of data enables consumers to rapidly and effortlessly access specific information and thereby reinforces the image of the destination and encourages people to visit it. However, this process is not without risks for those who plan and manage tourist areas, as it can lead to changes in tourist behavior and have a direct impact on the social, economic, and cultural life of the locations visited.

Although some attention has been given to the role of AI in shaping perceptions and planning decisions, there is a lack of empirical research on the spatial dimension of LLM-generated recommendations and their correspondence—or divergence—from established tourist behaviors [3]. This article analyzes the spatial location of the recommendations made by the ChatGPT language model (Chat Generative Pre-trained Transformer), which is one of the most efficient and powerful chatbots for information searches [15], and compares the results obtained from it with the places visited by tourists in Barcelona. In this study, we used Instagram, a social network based on audiovisual content, to identify the locations most commonly visited by tourists in Barcelona. Many tourists use Instagram to share images of the places they have visited, so it can provide reliable data relating to the spatial distribution of tourists at a given destination [16].

This study addresses the following primary research question: “How are the tourist attractions recommended by ChatGPT spatially distributed”. In relation to this initial question, the article also examines a secondary research question: “What changes can be identified between the tourist attractions recommended by ChatGPT and the areas visited by tourists prior to ChatGPT?”. In exploring these questions, the article analyzes the spatial outputs generated by ChatGPT and examines whether they can diverge from established tourist flows, thereby assessing the potential impact of LLMs on tourist decision-making.

This investigation into the spatial effects of AI-based language models is part of a broader line of research on place representation and spatial modeling. To date, relatively few studies have examined the potential influence of LLMs on the spatial distribution of tourist flows [3,6]. Our approach seeks to improve the understanding of urban tourism. AI offers new opportunities to analyze and extract the characteristics of specific places [17]. By focusing on the spatial effects of LLMs, this study aims to expand our understanding of urban tourism dynamics and explore the broader implications of generative AI for the management and planning of urban areas. AI not only offers new ways of describing places but also raises critical questions about how digital representations may shape real-world spatial practices.

2. Literature Review

This literature review is divided into three related sections. The first section explores how tourists move and choose destinations. It shows how digital tools like search engines and geolocation affect their behavior. It also explains how social media and user-generated content shape decisions and mobility patterns. The second section focuses on LLMs, especially ChatGPT. It examines how these tools change how tourists plan and interact with destinations. It highlights benefits like personalization and efficiency. At the same time, it points out concerns such as misinformation, bias, and user experience issues. The third section evaluates ChatGPT as a travel planner. It uses recent studies to show its strengths and weaknesses. These include useful features but also risks like hallucinations, privacy concerns, and lower reliability. Together, the three sections show how technology, tourist behavior, and destination management are evolving in the digital age.

2.1. Spatial Distribution and the Influence of Search Engines on Visitor Behavior

The compilation of information about the spatial location of tourists was traditionally carried out using surveys or observation [18]. The subsequent emergence of new instruments that allowed tourists to be geolocated resulted in an increase in both the quantity and quality of the data that could be processed, although there remained the important limitation of how to engage participants in these studies [19]. The current widespread use of smartphones, tablets, and cameras equipped with GPS has made it possible to create digital footprints that are easy to use in studies that identify places that are popular with tourists and the routes that they follow.

This transition to the digital world has also been observed with sources of information, with digital media having partly replaced traditional channels [20]. From the 2000s onwards, changes in technology and communications associated with the Internet and mobile devices began to provide tourists with a large quantity of new information which has enabled them to organize their own trips without the need for either travel agencies or Destination Marketing Organizations. As a result, tourist mobility patterns have become more diverse and have increasingly flexible. This, in turn, has made detecting supply-side opportunities a greater challenge, particularly when they are not directly connected to demand-side preferences [21].

This change in sources of information has led to corresponding changes in the attractions visited. Features regarded as essential elements of a particular destination normally motivate tourist visits [22], with tourists being attracted by idealized images and icons, which they then visit and immortalize in photographs [23,24]. This sets in motion a spiral of representation through which iconic images of a particular destination are perpetuated over time. It can, however, also generate new consumption patterns, which it is possible to detect by observing the digital footprints provided by visitors [5].

Social networks have produced a transformation in the way that visitors currently plan and share their trips. A conceptual study highlighted the fact that social networks are now an essential source of information during the planning phase of a trip [25]. During this stage, the information obtained by future visitors may influence the perceptions that they have of their potential destinations and the decisions that they take.

Social networks can also, however, produce an overload of information due to the great quantity of data now available. Future tourists will find themselves needing to take numerous decisions, which could result in mental fatigue [26], anxiety, and/or frustration, with them perhaps finally even failing to travel at all [27]. In this context, LLMs, including ChatGPT, can simplify the task of searching, taking the tension out of choosing between seemingly countless options [21].

2.2. Tourism and the Emerging Role of ChatGPT

Although academic research based on LLMs has grown swiftly [28], most studies have tended to focus on the following: their potential future uses in organizations’ internal processes and procedures; organizational networks and distribution; stakeholders; and customer preference services [3]. However, despite the use of these models by tourists may have a significant effect on tourist destinations, empirical studies analyzing the specific impact of LLMs on urban areas currently remain rather limited [15]. A situation that aligns with the way the tourism industry has incorporated LLMs. The travel and tourism industries have been using chatbots, robot staff, and self-service information points at least since 2020 [29]. However, until mid-2022, use of this technology was mainly confined to businesses. The arrival of ChatGPT marked an important leap forward as it opened up the use of LLMs to a much larger segment of the population. As a result, LLMs have now become an important target for tourism research [30].

ChatGPT has provoked a significant change, as it is capable of rapidly synthesizing information obtained from numerous sources and providing responses to questions. Unlike previous search engines, it has the ability to respond to users in natural language and to hold conversations with them. It cannot, however, differentiate between different types of sources [31]. The training data used by ChatGPT is fixed at a certain point in time and the model does not automatically update with new data [32]. Some of its responses may consequently be sub-standard as a result of systemic bias in the training data [33].

ChatGPT was developed by the company OpenAI and first unveiled in November 2022. As of September 2024, a free version—GPT-3.5—has been available, which is based on information from prior to January 2022, and a paid version—GPT-4—which uses data compiled before April 2023 also exists. The arrival of ChatGPT brought about a significant change as it means that AI can now be accessed by the general public [34], rather than only technical experts. This tool can assist users in trip planning, locating information about tourist destinations, and evaluating various alternatives. It also provides recommendations, which it does in a very personalized way, taking into account the interests and preferences that the user supplies to the language model. These different features have led several tourism companies to invest in this technology, as it can reduce their costs, improve their competitivity and the experiences of their guests, and also provide added value to its users [15].

ChatGPT has the potential to revolutionize business, tourist, and DMO (Destination Management Organization) processes in the tourism industry [35]. For tourists, one key feature is the “hyper-personalization” of marketing efforts [36]: it has the ability to generate personalized recommendations based on the specific preferences and patterns of each individual tourist. This can be exploited during all stages of a visit. During the preparation phase, it can be used to gather information and plan itineraries. During the purchasing process, it can help to verify aspects such as special needs, preferences, and budgets. Then, during the trip itself, it can be used to adapt the journey, schedule activities, and calculate routes. Finally, after the trip, it can help tourists to recall some of the places that they visited or even to generate reviews for social media.

On the other hand, the use of LLMs such ChatGPT can have some negative repercussions, one of which is generating fake reviews of hospitality and tourism organizations [34]. They can also have a negative impact on the tourist’s visit experience. The lack of technical competence of tourists using ChatGPT may result in them obtaining biased responses that do not correspond to the reality of the place that they wish to visit [35]. In addition, it could lead to a lack of social contact at the destination, which would have an adverse effect on the quality of the tourism experience [37]. There is also the risk of potential cybersecurity threats that could affect the reputation of the tourism company and/or destination [35]. All of these issues could potentially result in users losing trust in the service provided [38].

2.3. Strong and Weak Points of ChatGPT as a Travel Itinerary Planner

ChatGPT has been increasingly adopted as a planning tool in the fields of travel and tourism. Recent studies have shown that more and more tourists are now planning their future trips using ChatGPT [39]. Rather, [40] has presented empirical evidence that ChatGPT efficiently answers the questions posed by tourists, helping to improve their experiences and levels of satisfaction. Some authors even suggest that LLMs can optimize recreational offerings in cities (e.g., concerts, workshops, city tours) to attract more visitors [41]. Uses of ChatGPT in the field of tourism range from personalized systems and recommendations to robots, speech systems, intelligent travel agencies, and systems for forecasting and making predictions [42]. It is not, however, only companies that can benefit from the use of LLMs. Authors like Wong et al. [43] have shown how LLMs can also be used by tourists themselves, whether in pre-travel stages, during their trips, or after they have traveled. Such use could end up directly influencing tourist behavior at the final destination, either as a result of personalized experiences obtained from queries [39] or due to precise and rapid responses to queries relating to specific tourist products [44].

ChatGPT offers tourists strong points, such as its ease of use, which have favored its acceptance by part of the population [15]. It also provides personalized information and suggestions, which simplify trip planning for tourists and significantly improve the customer experience [34]. Alternatively, it can offer simple solutions for the transient needs, petitions, and preferences of tourists [35]. In short, it helps to retain potential clients who might not make a certain visit if they were to consult other sources [45].

On the other hand, ChatGPT could also provide tourists with incorrect, or entirely false, information [39]. This may be performed in a rather convincing way, which could cause confusion for tourists. “Hallucination” is the term commonly used when models generate false or misleading information [46]. What is more, the lack of any kind of reliable control mechanism, with which to identify a potential lack of truthfulness in an answer, only exacerbates the problem. This potential lack of quality in the information obtained may give the tourist a feeling of mistrust [47].

Several studies that have exclusively examined ChatGPT have likewise reported issues concerning the accuracy of the results produced, deficiencies in its capacity for critical thinking, and worries regarding the privacy and security of the user data [44]. Other authors have also raised alarm relating to the potential misuse of ChatGPT and how this could have a negative impact on the environment [15]. Another worry relates to deficiencies in the responses obtained due to the difficulties that LLM may have in understanding the context of certain conversations, the lack of transparency of the sources consulted by the model, or the lack of rationality or moral compass in the replies [45]. When future tourists discover errors made by ChatGPT, their perception of the tool may be adversely affected, thereby reducing their confidence in the system as a whole [15]. This is something that may eventually condition the adoption of this technology [35]. Studies such as that conducted by Ali et al. [44] have shown that some people remain skeptical of recommendations made by LLMs, as they perceive potential risks or feel uncomfortable when using it.

This article’s primary contribution lies in addressing these currently underexplored dimension in tourism research: the spatial implications of large language models (LLMs), particularly ChatGPT. By building on existing literature in place representation and spatial modeling, this study advances understanding of how LLMs may influence the geo-graphic distribution of tourist activity. While prior research has addressed various applications of LLMs in tourism, their potential to shape spatial behaviors through AI-generated representations remains largely overlooked [3,6]. By foregrounding these spatial effects, the study offers critical insights into the capacity of LLMs to reshape tourist flows, with broader implications for urban planning and destination management.

3. Methodology

This study uses Barcelona as a case study to explore the spatial effects of LLMs on tourist behavior. Section 3.1 outlines the city’s tourism profile, highlighting its status as a high-volume urban destination. Section 3.2 introduces the analysis conducted using geolocated Instagram images from August 2017 to capture tourist patterns prior to the widespread adoption of LLMs. Section 3.3 details the process of collecting and coding ChatGPT-generated recommendations in 2024, based on simulated multilingual queries. Finally, Section 3.4 describes the spatial analysis methods used to compare both datasets, employing GIS techniques to identify statistically significant clusters of tourist activity.

A quantitative research approach is adopted, relying primarily on spatial and statistical analysis of large-scale datasets to examine how tourist behavior in Barcelona is shaped by both traditional user-generated content (UGC) and AI-generated recommendations. A case study design was selected to allow for an in-depth spatial examination of a single, high-volume urban tourist destination. The choice of a quantitative, mixed-source design was driven by the need for comparability, replicability, and spatial precision. The combination of GIS-based methods—such as Inverse Distance Weighting (IDW) and Getis-Ord Gi*—enabled the detection of statistically significant clusters and the spatial interpolation of both datasets. This design was deemed appropriate for identifying shifting patterns in tourist concentration.

3.1. Case Study

The study area selected was Barcelona, a city which is located in Catalonia, in northeastern Spain. According to the Barcelona Tourism Observatory [48], Barcelona welcomed 15.6 million tourists in 2023, resulting in nearly 36.4 million overnight stays. The city has 457 hotels, which offer 76,662 beds. Barcelona also has a significant amount of private housing that is available for tourist use: 9818 housing units with 58,124 beds. On average, tourists stayed in Barcelona for approximately 2.6 nights in hotels and for 3 nights in private housing. With regard to the demographic profile of visitors, 65% were male. The average visitor to the city was 35 years old. Most of these tourists (51%) were visiting Barcelona for the first time, and 38% traveled with their partners. Leisure was the primary reason for travel for 69% of those visiting the city. The average daily expenditure per person was €141 (excluding transport and tourist packages). Regarding the origins of tourists staying in hotels, 19% were Spanish (with 40% of these being from Catalonia), and 12% came from the USA. The UK and France each accounted for 7% of visiting tourists, while Italy contributed 6% and Germany 5%.

3.2. Extraction of Data from Instagram

To analyze tourist behavior prior to the emergence of large language models (LLMs), images posted on Instagram during August 2017 were used. Three different days were selected: August 13 (a Sunday); August 16 (a Wednesday); and August 31 (a Thursday, which was also a rainy day). These dates were strategically chosen to observe variations in tourist activity on different days of the week and under different weather conditions.

It is important to bear in mind that our research was based solely on the analysis of publicly available Instagram images. Although the spatial distribution of publicly shared images can provide a valuable first insight into the tourist dynamics of a specific area, such images only partially align with the total number of photos taken by tourists at holiday destinations [49].

The research was based on searches for images classified as “Barcelona” by Instagram’s geographic location tool. We applied a predefined set of criteria in order to code the Instagram images and maintain the objectivity of the study. We implemented a structured coding process to ensure the reliability of our sample [10]. Two researchers worked together to categorize 0.5% of the total dataset using a single classification framework. After refining some of the coding parameters, the rate of agreement between the coders exceeded 93%.

While this manual approach was time-intensive, it offered greater accuracy than employing automated data extraction methods. Instagram’s built-in geolocation tool employs a predefined database of locations, which can sometimes lead to inaccuracies when trying to pinpoint exact sites. As a result, the automatically assigned locations associated with social media posts are not always completely reliable [50].

In total, 29,645 unique images were collected; we also registered the number of likes that each image received within 24 h of being posted. The ‘likes’ were utilized to gauge the degree of social engagement. The locations of images posted by the same users in the previous week were also examined in order to enable us to discard any images that may have been uploaded by local residents. The average length of a tourist stay in Barcelona is less than 3 days. Thus, when any location obtained a week earlier was also in Barcelona, the corresponding image was excluded under the assumption that the user was a local resident. When the identity of the photographer was not clear, the image was also excluded. Applying this approach, we determined that 12,435 photographs (42.4% of the total number of collected images) had been taken by tourists.

We employed the criteria used by Pritchard and Morgan [51] to map the data: any image in which over 50% of the space was occupied by irrelevant spatial elements, such as selfies, objects or sunsets, was excluded. The images were then geolocated using personal expertise and the Google Images App to search for similar images and find links to pages that could identify the different locations.

Finally, 10,590 of the 12,453 user-generated images were pinpointed. This corresponded to 85.2% of the images associated with specific locations posted on Instagram by tourists, with a margin of error of 0.37% and a confidence level of 95%.

3.3. Extraction of Data from ChatGPT

The data input from ChatGPT were obtained from a combination of results provided by GPT version 3.5 and GPT version 4. We did not observe any significant differences between the locations of the tourist attractions recommended by the two different versions of the App. To obtain our data, simulated prompts were entered via various user profiles and from a range of Internet Protocol addresses (IPs). The research was carried out in April 2024.

To design the prompt, the key question was characterized using information relating to the date of the visit, the average stay time, the type of accommodation used, and any other relevant expenditure, based on tourism statistics for the city of Barcelona [48]. In addition, we requested 50 places to visit. The decision to request 50 places was based on the ‘three days in Barcelona’ tourism proposal presented by Barcelona Turisme (the official tourism office of Barcelona). On this website, this number of visits is suggested for a three-day stay in the city [52].

The prompt entered was the same in all cases: information was requested about places to visit, and a series of characteristics were entered that were similar to those sought by a typical visitor to the city of Barcelona.

“Good morning. My partner and I would like to visit Barcelona during the month of August. Our idea is to stay for 3 nights in a four-star hotel. The trip is for leisure, and it’s our first visit to the city. We aim to spend around €72 per person per day during the stay (excluding the cost of the hotel). Could you list 50 places that we could visit, providing just the name and the location, ordering them from the most to the least interesting? Thank you.

Following Shin and Kang [32], the researchers posted identical prompts in new chats to avoid any of the responses being influenced by previous prompts. The responses obtained were very similar.

The queries were made in the native languages of a typical range of visitors to the city: 12% in Catalan, 19% in Spanish, 22% in English, 10% in French, 8% in German, 6% in Italian, 5% in Dutch, 2% in Portuguese and Polish, and 1% each in Swedish, Chinese, Danish, Korean, Norwegian, Hebrew, Romanian, Greek, Japanese, Czech, Hindi, Hungarian, Arabic, and Turkish.

Data collection was carried out until theoretical saturation was reached, that is, when a new query no longer yielded any additional tourist sites for the sample. A total of 10,000 responses were collected and coded by two different coders to ensure that the locations really existed. This resulted in 166 locations (1.7%) being discarded. In most cases, these discards were located just outside the city limits of Barcelona. In a few cases, they were either invented (38) or their references were too vague (e.g., “their cinemas,” 16 cases).

The coding was based on the name of the heritage element. As in the case of the Instagram images, a predefined set of criteria was applied to code the locations. Two researchers collaboratively categorized 0.5% of the total dataset using a unified classification framework. The only issue identified during this process was that ChatGPT occasionally used different names to refer to the same place: for example, “Camp Nou” and “FC Barcelona Stadium”. In such cases, both results were mapped.

3.4. Data Processing

Our analysis employed traditional hexagonal tessellations [53]. Following García-Palomares et al. [54], we used hexagons with side-lengths of 200 m and ArcGIS Pro 3.2 software for making calculations. A side-length of 200 m is considered to provide an optimal balance between the spatial resolution necessary for detailed analysis of the study area and the overall readability and interpretability of the resulting map. The number of ‘likes’ served as the weight variable for Instagram, whereas the number of mentions was used for ChatGPT in the hexagonal tessellations.

We used two different tools to identify spatial distribution patterns: Inverse Distance Weighting (IDW) and Getis-Ord Gi* (Hot Spot Analysis). To perform the various calculations required, the information was standardized to a base of 1. The IDW technique is a means of interpolation that is frequently utilized in Geographic Information Systems (GIS) [55]. IDW interpolation employs a linearly weighted blend of sample points to determine cell values, with the weight being determined by the inverse distance function. Like other interpolation techniques, IDW utilizes a series of sample points (L1, L2, … Ln) to determine the value obtained for a given location, L. This method assumes that the influence of the variable being mapped decreases as the distance from the sampled location increases. We also used a number of default values, while for areas lacking identified images, interpolations were based on the 12 closest spatial values. In order to compare our results, the calculations of the various values were performed separately. We then made a spatial join to calculate the differences. The Getis-Ord Gi* measures the degree of clustering of either high or low values. It starts with a z-value and uses the respective p-value to indicate the spatial agglomeration of values with statistical significance. The p-value obtained was less than 0.01 for all of the indicators considered. This enabled us to reject the null hypothesis that the phenomenon analyzed was randomly distributed.

Prior studies had suggested that tourist flows were influenced not only by local characteristics but also by those of neighboring areas [56,57]. However, one of the strengths of the Getis-Ord Gi* is its capacity to identify clusters as distance increases, without the cases with the highest values affecting the other results. We regarded a high z-value and a low p-value as being indicative of a spatial concentration of high values [58]. For an area to be considered a statistically significant focal point, it, therefore, needed to exhibit a high z-value and to be surrounded by other areas with similarly high z-values. Given the characteristics of the polygons (their size and distribution), Contiguity Edges Corners was considered the most appropriate method for defining the neighborhood. False Discovery Rate (FDR) correction was also applied. FDR potentially reduces the critical p-value in order to account for multiple testing and spatial dependency. It also estimates the number of false positives and adjusts the critical p-value accordingly, offering better performance than assuming each local test in isolation or employing traditional multiple test methods [59].

4. Findings

4.1. Image Location

The 10,590 valid mentions found on Instagram (Figure 1) were compared with the 9834 obtained from ChatGPT (Figure 2). This showed a significant difference in the number of locations identified: 1321 on Instagram as opposed to 215 on ChatGPT—six times fewer. Both sources highlighted a significant number of locations in the historic center of Barcelona and in the southern part of the city (the park and the museum area of Montjuïc). However, the Instagram distribution was clearly more extensive and also included attractions located on the outskirts of the city and on the coast (to the east).

Both sources reported a high concentration of the most mentioned elements. The 10 most cited places accounted for 35.9% of ChatGPT mentions and 34.7% of Instagram likes. However, only three elements (three Art Nouveau buildings: the Sagrada Família cathedral, Casa Batlló, and Park Güell), were included in the 10 most mentioned locations for both sources, with a relatively low degree of overlap. In contrast, there was a slightly more balanced distribution among the bottom 10% of least mentioned places, which accounted for 0.38% of ChatGPT mentions and 1.25% of Instagram likes. Even so, the distribution of attractions listed by both sources was in line with the concept of the long tail: a few tourist landmarks tended to attract the most interest, while the majority of points of interest attracted only a few people [22].

The concentration of ChatGPT images could be seen more clearly when we compared the results presented using hexagons (Figure 3). Of the 326 hexagons containing images derived from either Instagram or ChatGPT data, it was only possible to observe a balanced distribution (deviation of +/− 10% of the weight) in 114 (35.0%). In nine hexagons (2.7%), the ChatGPT images showed a significantly greater presence than those from Instagram (+25% mentions). In another 12 hexagons (3.7%), ChatGPT revealed a slightly higher presence than Instagram. In contrast, in 152 hexagons (46.6%), Instagram exhibited the greatest presence, while in 39 (12.0%) it showed a clear majority presence.

This spatial distribution highlighted some very distinct characteristics. The points with a balanced presence in both types of media tended to fall into one of two categories: they were either very frequently visited or rarely visited. The most visited attractions included the historic center of Barcelona, the Eixample (Catalan Art Nouveau buildings), and the Gràcia neighborhood. All of these locations had a strong presence on Instagram and were also widely mentioned by ChatGPT. The least-visited places included a significant number of locations with a very limited presence on both ChatGPT and Instagram. These were more specialized destinations that received fewer mentions on both Instagram and ChatGPT. Both sources, therefore, provided similar results at both extremes.

In contrast, in the locations where ChatGPT mentions predominated, we mostly found areas that attracted a large number of tourists. These included attractions in the historic center, such as Les Rambles, Santa Maria del Mar, the Sagrada Família, and Park Güell. Only occasionally were less well-known attractions, such as the Horta maze (in the north of the city) or the Palau Reial (in the west), mentioned. ChatGPT tended to highlight major tourist spots, which were also very crowded locations. These were also identified in the Instagram data but generally appeared with less intensity than on ChatGPT.

There were many more points of interest cited on Instagram than on ChatGPT. Examples of this could be found in most of the beach area and in the former industrial district of Poblenou (east). This suggests that tourists currently visit many more points of interest than those promoted by ChatGPT and that this medium tends to promote images that only capture a select few of the locations that attract tourists.

4.2. Clusters on the Getis-Ord Gi*

The Getis-Ord Gi* local statistic makes it possible to locate hot and cold spots within a given territory. Hot and cold spots are significant points with unusually high or low concentrations of data values relating to the phenomenon being analyzed. However, not all extreme values, whether high or low, are automatically identified. For a hot spot to be considered statistically significant, it must not only have a high value but also be surrounded by similarly high values. Conversely, a cold spot is a cluster of consistently low data values. In both scenarios, the local sum must significantly deviate from that of surrounding entities (calculated by aggregating the values of neighboring hexagons). This method enables the identification of areas that exhibit types of behavior that are different from those observed in nearby regions. In our case, these were linked to the distribution of Instagram and ChatGPT images.

The first phenomenon to consider was that no cold spots were detected for either our Instagram (Figure 4) or ChatGPT (Figure 5) data. There were, therefore, no areas of the city in which clearly lower presences of Instagram or ChatGPT images were detected with respect to the average value.

In contrast, several hot spots were identified. These were often very clear, although there were noticeable differences between the two sources. For example, 26 out of 313 hexagons containing images (8.3%) fell into this category for Instagram, while for ChatGPT, there were only 7 out of 126 (5.6%). The locations of these points further reinforced the impression provided by the absolute data: the points mentioned by both sources were spatially concentrated, but this was much more evident in the case of the ChatGPT images. The mentions on ChatGPT tended to be most concentrated in a very small part of the historic center: the Gothic Quarter, an area that contains several museums and iconic medieval buildings. There were concentrations of images on Instagram in all the areas indicated by ChatGPT, but there were also others in neighboring areas, such as the Eixample (the Catalan Art Nouveau district of the city), the north of the city, and la Barceloneta (the seaside district in the south-east).

The analysis of Instagram images revealed four hexagons (1.6%) with a confidence level of 95% and one (0.3%) with a level of 90%, while on ChatGPT there were two at 95% (1.6%) and one at 90% (0.8%). In all cases, these were areas near the most visited attractions, revealing the existence of transition zones. These were areas that contained significant volumes of mentioned attractions that did not display the highest values. They did, however, identify areas that were progressively tending to resemble those that produced confidence values of 99%.

5. Discussion

It appears that ChatGPT tends to recommend a very small number of areas, which are also popular tourist destinations, while Instagram tends to present a richer and more diverse image of the city. This prompts us to reflect on the role that ChatGPT may play in the future of tourism in cities. In line with the long tail concept [60], it could be argued that ChatGPT only recommended areas that are firmly consolidated in the collective imagination of Barcelona. In contrast, the image perceived by tourists on Instagram was richer.

In this regard, if tourists were to stop looking for information on platforms like Instagram and to solely focus on ChatGPT, there would be a risk of them receiving an impoverished image of the city and being directed to only a limited number of locations. This is particularly important when we consider that several authors [57] (p. 9), have already pointed out that “Instagram tends to diffuse an image based on central areas (those with the most tourism tradition) while relatively neglecting those of more peripheral spaces.

Identifying the potential effects of using LLMs is a subject that has already been explored [61]. In this regard, the present research highlights how the spatial concentration of tourist attractions on ChatGPT is clearly even higher than on Instagram. This is a phenomenon that could have significant implications for the management of tourist destinations. Visitor management involves actions aimed at influencing visitor behavior in such a way that overcrowding occurs less frequently [62]. In a context in which various destinations are considering diversifying the number of areas visited as a strategy for preventing overcrowding, models like ChatGPT may generate a contrary movement: reinforcing certain central destinations and making it harder for lesser-known destinations to gain visibility. In this regard, the spatial concentration of the locations cited by ChatGPT is very evident. While Instagram images tend to be spread across wide areas of a city, the attractions recommended by ChatGPT are generally concentrated in only a few central locations. This implies that beyond the repercussions of ChatGPT on the tourism industry that have been noted by authors such as Carvalho and Ivanov [38], we must also consider and discuss the territorial effects of ChatGPT.

Also related to this issue is the limited variety of destinations cited by ChatGPT. Certain Instagram users include less common tourist attractions in their images. These may be attractions that the Instagram community values as a demonstration of knowledge of the destination and the discovery of new attractions. In contrast, ChatGPT tends to repeat the same attractions that have already been identified, doing so more frequently. This could encourage tourist destinations to give less priority to innovation and to solely focus on promoting already established sites and experiences. However, the use of ChatGPT may also, in the medium term, enable a new approach to tourism management by applying “smart solutions” [63] or improving resilience through “real-time responses” [64].

Another question to consider is that ChatGPT tends to rely on outdated information. The model’s knowledge is fundamentally based on data stored up to a specific point in time. As Shi et al. [52] have observed, this limitation may delay the inclusion of new tourist products; it favors well-established ones, while more innovative products take longer to be incorporated into the tourist experience.

It seems that Instagram can be influenced by trends and fads, which, in a relatively short time, may drive new tourists to less conventional areas, potentially causing management, risk, and/or sustainability-related challenges for tourist destinations. In contrast, ChatGPT’s current configuration provides information that is less prone to such fluctuations. ChatGPT, therefore, poses a less problematical challenge for the management of tourist flows than Instagram. Nonetheless, ChatGPT provides results that simplify the information available on other platforms and creates a sense of hyper-personalization. For this reason, ChatGPT is rated more highly than passive applications [65]. In the medium term, ChatGPT has the potential to contribute significantly to enhancing the sustainability of tourism activities. Authors such as Sigala et al. [66] have shown how ChatGPT can help to rationalize the energy costs of certain tourism activities, support destination governance, and improve emergency information. The same authors have also pointed out that ChatGPT can help more tourists to access niche attractions; this could potentially promote the more balanced and sustainable development of tourist destinations. However, the data obtained to date show that ChatGPT has significant limitations when it comes to tailoring recommendations to meet individual tourist profiles. Instagram influencers seem to provide information that is more closely aligned with user preferences, while ChatGPT tends to provide less specialized data. For tourists with highly specific needs, ChatGPT may fail to highlight certain elements that are present at a potential destination. It consequently seems that, at least in its current configuration, ChatGPT is, therefore, not a tool that can help to mitigate over-tourism or to distribute tourist flows more evenly. These findings highlight how place images are constructed and perceived differently by various groups, reflecting the complexity and multiplicity of urban experiences [13]. On Instagram, this Place image is manifested through the contributions of numerous users, whose perceptions are partially shaped by the underlying Place identity [24]. In contrast, ChatGPT presents a perspective more closely aligned with Place reputation, understood as a consensual evaluation formed by external audiences that results in a dichotomous judgment. If a location meets certain criteria, it is deemed “good” and included in ChatGPT’s responses; if not, it is disregarded. The outcome is a spatial distribution of urban attractions concentrated in a limited number of “good” urban areas, while the rest of the city remains absent from the results. Given that image construction is a dynamic and ongoing process, it cannot be ruled out that this form of place reputation may eventually influence both place image and place identity.

One clear example of this phenomenon is ChatGPT’s difficulty in incorporating aspects related to the everyday cultural life of different destinations. It tends to promote a heavily heritage-focused image of tourist sites, while finding it more difficult to highlight events that are more closely tied to local culture or to aspects of intangible heritage. It consequently presents an image that predominantly reinforces tourism associated with material heritage. This may be regarded as a positive development, as it contributes to mitigating the overcrowding of events associated with intangible cultural heritage. However, it also reinforces a vision of tourism linked to material heritage elements, which may distance itself from the wealth of heritage for which various institutions are currently advocating.

6. Conclusions

This study set out to investigate how tourist attractions recommended by ChatGPT are spatially distributed and to explore what this distribution reveals about potential shifts in urban tourism patterns. The findings suggest that the recommendations generated by ChatGPT tend to reinforce traditional, institutional representations of urban areas, echoing generic images historically promoted by tourism boards and travel companies (

Table 1. Summary of Key Findings and Implications.

Theme	Key Findings	Theoretical Implications	Practical Implications
Tourism Impact	Tourists may develop a narrow understanding of cities, focused primarily on central areas and long-established attractions.	Tourists tend to explore only a narrow segment of the urban environment they visit. Difficulty in introducing new tourism products to the market.	Social media has transformed tourists from passive consumers into active contributors to destination narratives [4]. However, with ChatGPT, this participatory role may diminish, reverting tourists to a more passive stance.
Spatial Implications	Reinforces dominant tourist circuits, potentially increasing overcrowding in popular areas and marginalizing lesser-known neighborhoods.	Urban planners and tourism managers may need to intervene to redistribute tourist flows and promote underrepresented areas.	Consistent with the observations of authors such as [3,6], LLMs may influence how short-term visitors perceive and interpret urban space
Cultural Consequences	Risks eroding local diversity and weakening the social and cultural identity of urban spaces.	Raises concerns about cultural homogenization and the loss of intangible heritage in tourism discourse.	As some authors have pointed out [67], ChatGPT favors dominant narratives, marginalizing alternative voices.
Tourist Experience	Reduces tourists’ role as co-creators of meaning and place; may lead to isolation and anxiety due to lack of human interaction.	Suggests a need to design hybrid experiences that balance AI guidance with opportunities for social interaction and discovery.	The lack of nuance in AI-generated recommendations can impact Place reputation [13]. This may hinder tourists’ ability to act as co-creators of new images and representations of the place—a role they actively fulfill on social media platforms [47].
Accuracy Concerns	1.7% of recommended attractions were non-existent, highlighting risks of misinformation.	Emphasizes the importance of verifying AI-generated content and developing mechanisms for quality control in tourism information systems.	AI-generated content can present convincingly inaccurate information (“hallucinations”), which may erode user trust in the absence of effective verification mechanisms [39,46].

Source: own elaboration.

). As such, the model appears to prioritize mainstream and widely recognized landmarks, which may give users an incomplete or skewed view of the full range of experiences a city can offer. This reflects a partial fulfillment of the research question: while ChatGPT is effective at identifying major tourist sites, it appears to overlook more localized or emergent attractions, limiting its potential to reflect evolving tourist behaviors.

From a spatial perspective, this trend toward simplification raises critical concerns for urban destinations. If LLM-generated content continues to reinforce dominant tourist circuits, it may contribute to overcrowding in already saturated areas, while marginalizing less-visited neighborhoods or attractions. This not only affects the distribution of tourist flows but also has implications for the social and cultural identity of places, potentially eroding local diversity and making it more difficult to shift or renew destination imagery [67]. Moreover, the lack of nuance in AI-generated recommendations may hinder tourists’ capacity to act as co-creators of meaning and place representation—a role long acknowledged in tourism scholarship [47].

In general terms, the findings of this study align with research suggesting a metamodern shift in geographical thought [68]. LLM-generated recommendations oscillate between modernist centralization—through standardized outputs such as heritage-focused itineraries—and postmodern fragmentation, characterized by hyper-personalization often disconnected from local context. This oscillatory behavior reflects the nature of contemporary tourist conduct, which is no longer shaped solely by stable preferences but increasingly by shifting desires influenced by large language models. In this scenario, LLM outputs do not determine tourist flows outright but rather modulate them within a narrow bandwidth of predictability. These dynamics are particularly relevant in a context marked by a growing disconnection between the subject and material reality—a condition that may lead to epistemological uncertainty, relativism, or even nihilism [69]. As such, urban planning must account for the coexistence of certainty and uncertainty as a structural condition of tourism practice in the age of artificial intelligence.

However, as LLMs develop, they could give rise to a number of other scenarios. Some authors have warned that the widespread use of these models could reduce the social value of travel, potentially leading to tourists becoming increasingly isolated and guided towards custom-made routes [21]. A lack of human contact and social support at the destination may even induce loneliness and anxiety [70]. There is certainly a need for future studies to explore such questions.

This study reveals how the use of AI in the tourism sector could have effects that reach beyond controlling the internal processes and procedures of travel organizations. As authors like Gao [17] have pointed out, AI can be used to understand questions associated with human activities, but the results obtained to date have shown how AI can also contribute to misinformation: in our study 1.7% of the results obtained corresponded to features that did not, in fact, exist.

Limitations

Some of the limitations of the study derive from the relative novelty of the field. Firstly, only one language model (GPT-3.5) was used; others might have yielded different results. Furthermore, Barcelona is a city for which there is a significant amount of available information on the internet, which will have been used to train the LLMs. The results for other less tourist-oriented locations or areas, with less information available online, could be different. It should also be noted that LLMs are capable of interacting with prior prompts. Studies such as that by Tosyal et al. [2], which had participants interacting with models for between 5 and 10 min, may yield different results. Similarly, attention must be paid to the general level of acceptance of LLMs (and the type of audience that will use them). It is possible that only a relatively small proportion of travelers would trust them. A third aspect to consider is that the present study may yield different results in regions with technological disparities. The development of AI depends on factors such as infrastructure, the availability of high-speed broadband Internet connectivity, qualified IT professionals, and existing public policies [41]. These elements may vary significantly across cities, potentially leading to divergent outcomes. Future research should examine how the deployment and expansion of AI influence the indicators analyzed. Finally, since it is not currently possible to know what underlies the algorithms behind the recommendations made by most of the models, one important aspect to analyze in future research should be the relationships between various tourist destinations and those who own the models. Authors such as Carvalho and Ivanov [35] have pointed out that, in the future, LLMs may incorporate features linked to the values and/or marketing strategies of certain specific organizations and/or territories. Will tourist destinations be able to influence these results? If so, how? And what consequences will this have? All these issues may well have an impact on the image and future of tourist destinations.