Skip to Content
MDPIMDPI

Advanced

Administrative SciencesAdministrative Sciences
PDF
  • Article
  • Open Access

26 May 2026

Assumptions and Undeclared Selection Criteria: The Usefulness of Generative AI as a Travel Recommender System

Gulbali Institute, Charles Sturt University, P.O. Box 789, Albury, NSW 2640, Australia
Adm. Sci.2026, 16(6), 252;https://doi.org/10.3390/admsci16060252
Version Notes
Order Reprints

Abstract

This paper examines the trustworthiness of generative AI as a tourism recommender system by analyzing how ChatGPT5.2 responds to an open-ended, zero-shot prompt: “Recommend me a list of 10 German Christmas Markets.” Using German Christmas markets as a case study, outputs, texts in reasoning panels, and cited sources of fifteen replicates (carried out over five consecutive days) were systematically documented and analyzed. The results show a consistent and patterned selection which is dominated by a small canon of markets (Nürnberg, Dresden, Köln, München, and Stuttgart). The generative AI model does not neutrally sample from the entire pool of approximately 2000 German markets but instead reproduces a narrow canon of “iconic” destinations. Analysis of reasoning traces and follow-up conversations demonstrates that ChatGPT5.2 applies hidden selection criteria, including canonical status, landmark setting, branding strength, and perceived trip-planning usefulness, while also introducing undisclosed filters such as geographic spread across Germany and stylistic diversity. Although the model claims to use source triangulation and quality checks, the evidence shows substantial reliance on tourism marketing pages, travel media, blogs, and social media, especially for descriptive commentary. The study concludes that generative AI tourism recommendations are useful but non-neutral and should be interpreted as “curated,” bias-bearing constructs rather than transparent information retrieval. The implications of this on tourism management and the marketing of Christmas markets are discussed.

1. Introduction

Ever since the release of GPT3.5 by OpenAI on 30 November 2022 (OpenAI, 2022), Artificial Intelligence (AI), and in particular generative AI, has been poised to become the greatest disruptor to knowledge management since the creation of the world wide web in 1993 (Spennemann, 2025b). While the full extent of its impact is still evolving, it has manifested itself as a social disruptor on a cross-sectoral and global scale and, at the time of writing, has pervaded almost all aspects of human life. For many, ChatGPT has become a “go to” tool to support decision making such as meal recipes based on available ingredients or travel suggestions.
Traditionally, recommender systems functioned as decision-support tools that helped travelers choose destinations, hotels, restaurants, or itineraries by filtering structured information. Three main filtering systems existed: content-based, knowledge-based and collaborative. Systems relying on content-based filtering matched user preferences such as budget, location or activity type with destination or service attributes. Knowledge-based filtering supported personalized trip planning by matching user preferences such as season, distance or travel purpose with possible itineraries. Systems deploying collaborative filtering recommended places liked by users of similar profiles. While some systems were standalone, others employed hybrid approaches (Borràs et al., 2014; Burke, 2002; Loh et al., 2003). These recommender systems were prone to biases by over-recommending well-known destinations at the expense of locations for which limited data were included in the databases. Mainstream preferences could be reinforced due to demographic and cultural biases both by the database managers and users (in collaborative filtering systems) (Alves et al., 2023; Banerjee et al., 2023). These systems were also prone to commercial interests by favoring some destinations and service providers over others. Users were often locked into selective “ecosystems” where destinations, activities, accommodation, gastronomy and transportation were interlinked (Almeida-Santana et al., 2020; Khan et al., 2026). In addition, traditional recommender platforms were constrained by platform visibility and search ranking (Smithson et al., 2011; Xiang et al., 2009).
Generative AI models, such as ChatGPT, have been embraced by the tourism industry not only in order to automate repetitive tasks or respond to consumer feedback and reviews (Dogru et al., 2025; Gursoy et al., 2023) but also to enhance strategic capabilities. These range from tourism providers being able to improve the segmentation and personalization of products and services to consumers (Florido-Benítez, 2024; Kshetri et al., 2024) to pricing optimization. It allows tourism organizations and providers to create tailored content in real time (Jung et al., 2026). Beyond this, generative AI is poised to redefine how tourism experiences are imagined, delivered, and evaluated (Christou et al., 2025; Nicolau, 2025).
The benefits of generative AI models are that complex tasks can be completed near instantaneously, in timeframes that are much shorter than humans are capable of. The negatives are that AI, in particular generative AI, functions as a black box where the steps of “reasoning” that lead to the output are not always evident or transparent.
While generative AI as a computer model might appear value neutral, numerous studies have shown that algorithmic biases exist that influence a generative AI model’s response. These biases are derived from the selection of data (documents, websites and images) used for training the models, as well as any subsequent moderation in the model training and then in the final testing (red teaming) phase. This affects all generative AI models, both large language models and text to image generators (Fujimoto & Takemoto, 2023; Gross, 2023; Jenny et al., 2023; Motoki et al., 2023; Nakashima et al., 2023; Rutinowski et al., 2023; Spennemann, 2025d, 2025f; Sullivan-Paul, 2023; Szymański et al., 2025; Thomson et al., 2025).
There is a considerable body of work that has examined the utility of ChatGPT as a recommender system (Dai et al., 2023; Marti et al., 2025). Studies have shown that including user demographic information in prompts will affect model biases and stereotypes (Deldjoo, 2025). A number of specialized recommender systems have been developed for applications in various industries and disciplines (Guo et al., 2025; Wang et al., 2025). The majority of the general public, however, will not seek such specialized systems and will treat ChatGPT as a “go to” source.
At present there is only a small body of literature that canvasses whether the travel information and advice provided by generative AI models are trustworthy for travel planning and booking. Most papers have looked at consumer perception (Arora et al., 2025; Foroughi et al., 2025; Kang et al., 2026; Koçak, 2026; Luo et al., 2025; Wong et al., 2025). What is lacking are studies that investigate the nature and quality of advice of the travel information provided by generative AI models. The question arises, therefore, of to what extent generative AI can serve as a recommender system in the tourism sector, to what degree algorithmic biases will shape the response and what administrators of tourism organizations and events can do to mitigate the impact and maximize the return offered by generative AI. Using German Christmas markets as a case example (see Section 2.5 for background), this paper will explore how ChatGPT selects tourism destinations when given a broad open-ended prompt. ChatGPT was chosen as it by far the most frequently used, freely accessible generative AI model with 900 million weekly active users (Singh, 2026). The rationale for using a broad open-ended prompt derives from the aim of simulating the query of a curious, but as yet uninformed, customer. The first set of data, in this case a seemingly ranked list of Christmas markets, will inevitably influence, if not shape, the user’s opinion, at least on a subconscious level, thus biasing future follow-up questions. While the rankings and ChatGPT5.2’s chosen attributes of the markets will be discussed, the main emphasis of this paper is on ChatGPT5.2’s selection process of the markets as listed, on the quality assurance processes espoused by the model, as well as on underlying biases that the consulted sources may generate.

2. Methodology

OpenAI’s ChatGPT5.2 was used as the generative AI engine. ChatGPT5.2 is a multilingual, multimodal generative pretrained transformer, which was released by OpenAI in December 2025 (OpenAI, 2025b, 2025d) and which was built on previous iterations of the GPT-5 model (OpenAI, 2025a, 2025c). Guided by the instructions and context in the user prompt, ChatGPT5.2 generates text by using a neural network trained on large datasets to estimate the probability of each possible next word given the preceding context, then selecting the most probable sequence to form a coherent response that is relevant and aligned with what the user has asked. ChatGPT5.2 was trained on publicly available texts such as websites, articles, forums, and other content that is accessible on the open internet; licensed datasets and corpora; and text written by human trainers used to teach the model how to follow instructions, such as: demonstrations of good answers and comparisons/rankings of responses (Altman, 2023a; OpenAI, 2023, 2025d). Some of the corpora are extensive collections of out-of-copyright texts (Knibbs, 2024a, 2024b). The training data cut-off for ChatGPT5.2 was August 2025 (Conversation A) (Spennemann, 2026c). The data accessible to ChatGPT5.2 are the same or better than those accessible to earlier models. This encompasses, inter alia, the prior mentioned corpora and all material that can be publicly crawled on the internet, which includes academic open-access papers hosted publicly by publishers as well as the repositories of universities. In addition, OpenAI has entered into non-disclosed proprietary data partnerships that include paywalled content, archives, and metadata (OpenAI, 2024). Internet data include Wikipedia (Altman, 2023a, 2023b; Spennemann, 2025e) as well as such social media material (e.g., Facebook pages) where the authors have made them publicly searchable (OpenAI, 2024).

2.1. Prompting

The prompt provided to ChatGPT5.2 was deliberately formulated in a zero-shot unconstrained fashion in order to ensure that the model’s responses were not biased towards the author’s predepositions and perceptions. The following prompt was used: “Recommend me a list of 10 German Christmas Markets.” As requested, ChatGPT5.2 returned a list of ten Christmas markets, adding a brief, one sentence explanation to each. That listing was preceded by four photographs of Christmas markets taken from the internet. The data for the study were generated on five consecutive days in sets of three replicates (“runs”) each (Table 1). The spread over five days aimed to negate any “memory effects” that may have emerged from repeated identical prompting over a short timeframe. The average time ChatGPT5.2 needed to complete a prompt task was 1:10 ± 0:20 min (median 1:06 min, range 0:36–2:07 min).
Table 1. Details of the data runs.

2.2. Data Curation

Each “conversation” as well as the “thought” sequences were copied into MS Word and saved in individual files. The chat was then deleted (with automatic archiving turned off) in order to ensure that each response run started from a clean slate and ChatGPT5.2 could not refer to prior chats. The individual “conversations” have been documented according to a protocol (Spennemann, 2023) and have been archived as a supplementary data document (Spennemann, 2026b). Extracted and tabulated data of sources (websites) and market descriptors have been archived as a supplementary MS Excel data document (Spennemann, 2026a).

2.3. Documenting ChatGPT5.2 Reasoning

A pop-out “reasoning panel” provides text of the “thought” sequence that ChatGPT5.2 followed to arrive at the response. This comprises action statements and a series of web links that were accessed to complete the task. Once the list of markets had been finalized, ChatGPT5.2 searched the web to gather information for the sentence explanation of each selected market. Links to these sources were appended to the end of the “thought sequence.” Although the full text of the reasoning panel was copied, the links to the websites in the reasoning did not copy. It was not noticed until run 9 had been completed that only the final list of sources appeared in the clipboard. To remedy this, the relevant “thought” sequence sections were screen-captured for runs 10 to 15. In addition, video recordings of the screen were taken during this “thinking process” to capture the progress as reported by ChatGPT5.2 and correlate this with the text reproduced in the reasoning panel. The combined reasoning texts for each run were saved in individual files and, in combined form, archived as a supplementary data document (Spennemann, 2026d).

2.4. Other “Conversations”

In addition, three “conversations” were carried out to illuminate the nature and time cut-off of the training dataset (Conversation A), the process as to how ChatGPT5.2 compiled the list of 10 German Christmas markets in response to the prompt (Conversation B) and how ChatGPT5.2 chose and evaluated the validity of the sources it drew on (Conversation C). These have also been documented and have been archived as a data document (Spennemann, 2026c). The citation convention espoused in this paper denotes the respective conversation, identified by a letter, and the prompt in the sequence, identified by a numeral (e.g., “Conv. B2”).

2.5. Background to the Case Study

In Germany and German-speaking communities in surrounding countries, Christmas markets (“Christkindlmarkt,” “Weihnachtsmarkt”) form an integral component of the weeks in the runup to Christmas (BDSM, 2001; Spennemann, 2025a), Many markets, mainly in the larger towns, have a long tradition, some reaching back over 400 years (e.g., Strietzelmarkt in Dresden since 1434) (Spennemann & Parker, 2021). Christmas markets are widely promoted as a quintessential German tourist destination during December, with over 2000 locations attracting large numbers of local, domestic and international visitors (Jürgens, 2014; Spennemann, 2025c). While the COVID-19 pandemic of 2020–2022 caused the cessation of all markets in 2020 and a slow recovery since then (Spennemann, 2025d; Spennemann & Parker, 2025), Christmas markets are being understood and marketed at intergenerational multisensory experiences (Spennemann et al., 2025).

2.6. Limitations

The study only uses one generative AI system, ChatGPT, and does not provide a comparative analysis of other models. This is deemed justified given that ChatGPT is by far the most commonly used freely available system (Singh, 2026). As it focusses on a single topic (German Christmas markets) the study functions as a case study. The fact that the data were collected in February, i.e., outside the season of the Christmas markets, does not impact the data as the Christmas market resources that ChatGPT drew on exist all year round.

2.7. Trustworthiness

For the purposes of this paper “trustworthiness” relates to the degree to which responses provided by ChatGPT can be relied upon as accurate, credible, and dependable for decision making. It involves not only the accuracy of the content and the quality, diversity, and credibility of the data sources used but also the extent to which the processes behind generating that content are transparent, explainable, and consistent.

3. Results

The fifteen individual runs returned a total of seventeen different Christmas markets. Five of the markets (Dresden, Köln, München, Nürnberg, Stuttgart) were included in all fifteen runs, with five others being included in two thirds of all runs (Aachen, Berlin Gendarmenmarkt, Heidelberg, Leipzig, Rothenburg o.d. Tauber) (Table 2). When considering the average rank attributed to the markets, the sequence of the top three (Nürnberg, Dresden, Köln) is uniform across all responses. The other two markets that are represented in each of the responses occupy ranks 4 (München, 4.33 ± 0.62) and 5 (Stuttgart, 5.00 ± 0.76) with Berlin Charlottenburg (rank 5) being omitted as a singleton (Table 2). In the third quartile (50–75%) are the markets of Berlin Gendarmenmarkt, Frankfurt and Heidelberg. The remainder are in the lowest quartile.
Table 2. Rankings of German Christmas Markets by representation in the lists and average rank.
Each of the fifteen responses was preceded by a “carousel” of four square images of Christmas markets which were taken from the internet (Spennemann, 2026e). Six markets were chosen by ChatGPT5.2, with imagery of Dresden and Nürnberg included in every response, followed by Köln in two thirds of the responses. Rothenburg o.d. Tauber, Berlin (Gendarmenmarkt) and München were represented in one third of the responses or less (Table 3). The sources of the images provided by ChatGPT5.2 are primarily Wikimedia Commons for freely-licensed photos (27.8%), “official” tourism bureau (destination marketing organization) sites (27.8%) and licensed commercial stock photo sites (Getty, DPA, Flickr) (22.2%) (Spennemann, 2026e).
Table 3. Representation of German Christmas Markets in the image carousel per run and the rank of the represented market.
The listings were provided without any stated reasons why each of the ten markets was chosen over others. What ChatGPT5.2 provided for each of the chosen markets was a brief descriptive commentary that often bordered on hyperbole. Examples are:
“Aachen—Super charming near the cathedral and town hall; great for a cozy weekend trip.”
(run 7)
“Dresden—Germany’s oldest/most historic-feeling big market, known for stollen traditions and a huge lineup of stalls.”
(run 6)
“Nuremberg—The classic, world-famous market in the main square with rows of wooden stalls and a strong ‘storybook’ vibe.”
(run 1)
“Rothenburg o.d. Tauber—Peak fairytale medieval setting; half-timbered houses, lantern glow, and a famously atmospheric layout.”
(run 5)

Data Sources

The data sources drawn by ChatGPT5.2 illuminate some of the background to the choices of markets as offered by ChatGPT5.2. These are represented in the “reasoning panel,” an expandable section in the message composer that records the model’s thought sequence. While the panel can be accessed during or after a response to a prompt, the content is erased for each prompt even though the main chat is preserved. There are two sets of data sources: those that were drawn on during the initial selection of markets (Table 4) and those that were consulted when developing the brief descriptive commentaries (Table 5).
Table 4. Data sources mentioned in the reasoning panel during the “thought” sequence (% per run).
Table 5. Final data sources listed in the reasoning panel (% per run).
Across runs 10 to 15, the process for the initial selection of markets drew on 246 mentions of 52 different websites (Spennemann, 2026a). The dataset is quite diverse without a single source dominating. The most frequently used site was the German newspaper Die Welt (welt.de, 6.9% of all sites) followed by christmasmarketsgermany.com (5.3%) and facebook.com (4.5%). All sites can be assigned to source classes which can be aggregated into major groups: official sites (e.g., market organizers, destination marketing organizations) (50.3%), (semi-)formal service sites (e.g., newspapers, Wikipedia) (30.1%), purely commercial providers (7.3%) and private sites and social media (7.3%) (Table 4).
When considering the sites that were consulted in order to develop the brief descriptive commentaries, the number of websites increased considerably to 1650 (Spennemann, 2026a). Again, the dataset is quite diverse without a single source dominating. The most frequently used site was Facebook (10.9% of all sites) followed by besteuropeanchristmasmarkets.com (4.7%) and instagram.com (4.5%). Most cited sites were only rarely drawn upon, with 98.3% of all 1650 sites being represented only once. When grouped into the four source classes, official sites are in the majority (35.9%), closely followed by private sites and social media (32.2%), while (semi-)formal service sites (21.0%) and purely commercial providers (10.7%) are less represented (Table 5).
When comparing the proportions of the source classes used for the initial selection with the sites that were consulted when developing the brief descriptive commentaries, significant differences emerge. The representation of official sites decreases significantly (χ2 = 11.916, df = 1, p = 0.0006) as does the representation of (semi-)formal service sites (χ2 = 6.525, df = 1, p = 0.0106) (MedCalc Software, 2026). Private and social media sites on the other hand were very significantly more consulted when developing the brief descriptive commentaries (χ2 = 39.697, df = 1, p < 0.0001). Even when comparing between the source classes of sites consulted when developing the brief descriptive commentaries, private and social media are significantly more represented than official sites (χ2 = 4.757, df = 1, p = 0.0292) and (semi-)formal service sites (χ2 = 53.874, df = 1, p < 0.0001).

4. Discussion

Given that annually there are approximately 2000 Christmas markets in Germany (Wintertraum GmbH, 2024), many of which attract millions of visitors each (Spennemann, 2025c), the question arises of how ChatGPT5.2 arrived at the ten markets that it provided in response to an unconstrained query. When posed as a prompt at the completion of data run 15, ChatGPT5.2 provided the following five steps (Conv. B1) (Spennemann, 2026c):
  • “Start with the “canonical” markets almost everyone agrees on;
  • Ensure geographic and “vibe” diversity (not 10 markets that feel the same);
  • Preference for markets with a clear identity and a stable “home base”;
  • “Reality check” each entry: does it have an official event/tourism listing?
  • Keep it to a “top 10” that’s useful for trip planning.”

4.1. What Defines a Christmas Market as “Canonical”?

“Canonical” markets were defined as those that “are consistently described as among Germany’s best-known and most visited” and “show up repeatedly in reputable roundups and travel guides.” To ensure diversity, the algorithm selected some “big-city, high-production markets… for accessibility, transit hubs, and scale,” some “storybook/historic old-town markets …for atmosphere” and some markets with “a distinctive “theme” market … so [that] the list is not all standard stalls and Glühwein” (Conv. B1).
ChatGPT5.2’s algorithm prioritized markets “with a clear identity,” such as being “strongly associated with a specific square [or] setting” and being “known for something specific [e.g.,] a signature food, craft tradition” (Conv. B1). In its selection criteria ChatGPT5.2 prioritized markets that have a “globally iconic ‘signature market brand’ that is unusually famous in Christmas-market culture” (Conv. B10) and that “most people would be disappointed to miss” as part of their bucket-lists (Conv. B11). These markets commonly exist in “a landmark setting that is almost always the headline image” (Conv. B10) and that gives the prospective customer a “predictable high impact atmosphere” in “photogenic, unmistakable settings so [that their] experience matches expectations” (Conv. B11). The quality control check included verification that each of the chosen markets had a “credible official page (city tourism board or the event’s official site).” The final selection of the “top ten” markets was based on the market being bucket-list famous, uniquely memorable, or an excellent anchor for a region or city trip (Conv. B1).
As the prompt had been purposefully written in an open-ended fashion (“Recommend me a list of 10 German Christmas Markets”), any selection logic and ranking reflect the algorithm underlying ChatGPT5.2. The model evaluates a market as being “canonical” if it reliably meets most of these criteria: “repeated inclusion across many independent reputable lists,” being “frequently described using reputation-defining labels,” possessing “strong historical continuity and cultural symbolism,” and is of a “large scale and/or [possesses a] landmark setting that amplifies prominence” (Conv. B2). “Markets that are repeatedly described (by institutional sources) as ‘oldest,’ etc., become default picks because history is an easy, stable reason for fame” (Conv. 12).
The first listed inclusion criterion, “the repeated inclusion across many independent reputable lists” aims to triangulate “across different types of sources (institutional pages and major travel media and reputable features)” (Conv. B12). This, then, goes to the heart of what sources ChatGPT5.2 is trained to consider authoritative. It repeatedly cited Condé Nast Traveler (cntraveler.com) and Time Out (timeout.com), but also “A Dangerous Business” (dangerous-business.com) and Archeology Travel (archaeology-travel.com) (Conv. B1, B2). Travel lists seem to be considered reputable if they are derived from “major travel outlets and long-running travel blogs/guides” (Conv. B2). ChatGPT5.2 bases its quality assessment on the notion that while such sites are not “primary sources…they are credible for consensus and comparisons because they have editorial processes, brand reputation, and consistent travel coverage.” It acknowledges, however, that travel review sites “can be subjective, trend-driven, and sometimes influenced by what is easiest to cover” (Conv. B4).
There is a body of literature that examines the nature and reputation of online travel reviews (Guzzo et al., 2022; Marchiori & Cantoni, 2011; O’Connor & Assaker, 2024), as well as the web presence of online travel agencies (Zhu et al., 2022). Some work has also been carried out on the reputation of travel blogs and vlogs and their authors primarily from the viewpoint of clients and followers (Mainolfi et al., 2022; Nguyen et al., 2025; Tan & Chang, 2011). There is neither an accepted methodology to assess veracity let alone bias of these sites, nor is there an academically curated list of websites with quality assurance processes that are deemed reputable. In consequence, there is no simple way to independently assess the level of quality assurance these “reputable lists” entailed.
When taking the travel sites at face value, the internal quality assurance processes are left wanting. Examining the Condé Nast Traveler (cntraveler.com) site, for example, review standards and criteria are clearly spelled out for hotels, restaurants, rentals and villas as well as cruise ships and travel gear, but destinations and experiences as such, as well as destination lists, are not covered by the standards (Condé Nast Traveler, 2026). Time Out uses experts to assess (“testing and tasting”) but does not publish a clear set of evaluation criteria (Calhoun, 2025). Archeology Travel is a travel review and travel brokering site that has published a code of ethics but fails to publish a clear set of evaluation criteria (Travel, 2023). “A Dangerous Business” is a personal travel blog reviewing and promoting places and experiences, which also does not publish a clear set of evaluation criteria (Williams, 2025). Given that clear sets of evaluation criteria are not available for three of the four key sources ChatGPT5.2 drew on and given that none of the criteria of the fourth source applies to destinations, doubts arise as to the underlying purported quality control.
The ranking among the identified sites appears to select “markets with clearly identifiable “flagship” branding/pages” (Conv. B9) and the use of “reputation-defining labels” therein (Conv. B2). Their trustworthiness as constructed by ChatGPT5.2 is based on the fact that “they are run by the organizers or closely affiliated institutions [who] have direct authority over logistics” and “official features” (Conv. B4). That text, however, is written by the marketing departments of the market organizers of the communal official destination marketing organizations and therefore includes selective wording, statements or claims that trigger an emotive response and embellish reality without purporting untruths (Spennemann et al., 2025).
Indeed, the response by ChatGPT5.2 with regard to accuracy and trustworthiness of sources formally acknowledges this by stating that ChatGPT5.2 will “treat [superlative claims of market organizers] as ‘facts’ if they are widely corroborated” and that destination marketing organizations may “emphasize highlights and downplay negatives” (Conv. B4). Yet in the same response this is relativized as one of its “internal rules of thumb” and that for the assessment of superlatives it “prefer[s] authoritative historical/official sources; otherwise [would] treat [them] as ‘widely claimed’ rather than proven” (Conv. B4). While claimed, that is not adhered to. For example, when asked to regenerate a list drawing only on “primary/institutional sources (official market websites and city/regional tourism boards)”, ChatGPT5.2 also drew on the German newspaper Die Welt (welt.de) as well as the German tabloid Bild (bild.de) (Conv. B8). The brief commentary of market characteristics that was provided for each market in that rerun also drew on sites such as Facebook, Instagram and YouTube even though only primary and institutional sources were to be used (Conv. B8). Again, this points to an algorithmic issue, where the initial reasoning to a user prompt follows the specific instructions of the prompt but where ChatGPT5.2 defaults to its generic approach thereafter.
While ChatGPT5.2 may not cite social media in the listings, they are included in the search pattern. As ChatGPT5.2 responded when specifically queried, social media “can influence the search path (what [ChatGPT5.2] checks next), even if they do not determine the final claims” (Conv. B14). ChatGPT5.2 further responded that “the presence of social media sources in the reasoning pane can still be a problem because: they can prime me toward the usual popular’ narrative, can exclude quieter but equally legitimate markets, and can sneak in common knowledge’ phrasing that is actually just repeated marketing” (Conv. B14).
When examining internet sources for market-specific details, ChatGPT5.2 selected sites to “balance three things: accuracy (primary info), trustworthiness (who’s saying it), and consensus (independent repetition)” (Conv. B3). This resulted in the following hierarchy: official market sites; city/regional tourism boards (for credible institutional confirmation); major travel outlets (for consensus and comparative context) and Wikipedia (to cross-check for widely cited historical facts). Largely excluded were affiliate-heavy listicles with no clear sourcing and obscure blogs with no evidence. At the same time, blogs that were assessed to be clearly written and specific, with information that “matches what higher-quality sources also say,” were included (Conv. B3).
Using Wikipedia to cross-check for widely cited historical facts and thus validate the choice of markets generates a circular argument, as Wikipedia pages are written by crowd sourcing, with authors drawing on the official tourism websites (market organizer or communal/regional destination marketing organization) as well as media sites for factual information. An assessment of 35 Wikipedia pages that described individual Christmas markets in Germany found that only 11.3% of all cited sources were formal publications of archival sources. The majority were newspaper items followed closely by webpages published by market organizers and destination marketing organizations. Even travel websites were cited as evidence (on 14.3% of all pages) (Spennemann, 2026f).

4.2. Selection of Geographical Spread

Even though ChatGPT5.2 was only prompted to “Recommend me a list of 10 German Christmas Markets,” the algorithm provided a listing with a geographic filter, ensuring a spread across the country. One of the five criteria applied by ChatGPT5.2 to the generation of the list was to “keep it to a ‘top 10’ that’s useful for trip planning” (Conv. B1). ChatGPT5.2 reasonably interpreted the formulation of “recommend me” in the unconstrained (zero-shot) prompt to offer up a “top 10” list. Yet the wording of the prompt does not imply that the user wished to engage in trip planning. That assumption, which then introduced a geographic element into the site selection, is an algorithmic bias. When prompted, ChatGPT5.2 acknowledged the intentional selection bias to generate a set of markets that spans Germany geographically (“one-per-region-ish coverage”) and covers different market styles (Conv. B9). That bias, however, is not made explicit in the responses.
Rather, the user is presented with a list that is preceded by a statement such as “Here are 10 of the most iconic German Christmas markets to put on your shortlist” (Run 6). The laudatory qualifier terms used by ChatGPT5.2 were “classic” (run 2), “standout” (12), “most loved” (7), “most iconic” (6), “top” (14) as well as “excellent” (3, 13). Six of the lists were introduced as “can’t-miss” markets (runs 1, 4, 5, 8, 9 10, 11). These formulations insinuate that all markets were treated equally in generating the list, rather than moderating them through geographic and stylistic lenses.
Furthermore, as noted, the algorithm provided a listing numbered from one to ten. Unlike bulleted lists, which still have a sequential ordering, numbered lists also generate a perception of ranking in the mind of the user, irrespective of whether such a ranking was intended.
ChatGPT5.2 concluded each listing with a helpful prompt (“If you tell me…”) asking the user to clarify on which dates the user wanted to travel (all runs), what “vibe” (e.g., “big-city energy or fairytale towns”) or offerings (e.g., food, handcrafted gifts, scenery) were desired (10), or which region the user wished to visit (four). It was suggested to condense the list to a “tight” or “easy” best-fit itinerary of 2–4 or 3–5 markets including easy train pairings (three runs).

5. Conclusions

This paper set out to examine to what extent a generative AI model, such as ChatGPT5.2, can serve as a recommender system for tourism destinations when faced with an open-ended, zero-shot prompt and to what degree algorithmic bias shapes that recommendation. The study could demonstrate that the model does not generate a neutral list of options from the pool of approximately 2000 German Christmas markets but instead generates a highly stable, internally patterned, and implicitly curated selection. The output of fifteen runs used as replicates was dominated by a small core of markets (Nürnberg, Dresden, Köln, München, Stuttgart). This consistency indicates that ChatGPT5.2 repeatedly reproduces and reinforces a small canon of markets it deems to be “iconic,” “classic,” or “can’t-miss” rather than sampling from a broad field of plausible destinations.
The analysis of the ChatGPT5.2 reasoning panel and the supplementary “conversations” demonstrates that this hierarchy is caused by a selection logic that favors “canonical” status, stable branding, landmark settings, and perceived usefulness for trip planning. This indicates ChatGPT5.2’s algorithmic preference for markets with strong reputational attributes, extensive web presence, and repeated mentions across travel and institutional pages. While ChatGPT5.2 formally states an internal quality assurance reasoning, based on triangulation, source authority, and consensus checking, the empirical evidence strongly suggests that these processes are only partially adhered to. The model’s reliance on travel media, blogs, and social media, as well as destination marketing content, especially for the descriptive commentaries, is of concern. It sets up conditions for “inherited” promotional bias to persist, for the reinforcement of preexisting narratives, and the reproduction of “common knowledge” phrasing derived from marketing pages.
ChatGPT5.2 also injects a trip-planning frame into an otherwise unconstrained prompt, including assumptions of geographic spread and stylistic diversity, without disclosing this filtering logic to the user. As a consequence, a user may interpret the returned numbered list as a ranked and comprehensive recommendation, even though it is in fact the product of hidden selection criteria. This has implications for trustworthiness of generative AI models in tourism settings. While ChatGPT5.2’s outputs are coherent and may appear to be useful and practically convenient, they fail to be transparent with regard to the inherent algorithmic biases. In consequence, tourism recommendations derived from generative AI should be interpreted as biased constructs rather than neutral information retrieval.

6. Implications for Tourism and Market Administrators

The following section outlines the implications of the study for regional tourism administrators, organizers of well-known, but not iconic, Christmas markets and organizers of iconic Christmas markets. While focused on Christmas markets, the underlying principles of the impacts apply to conceptually similar offerings and can be extended to regional tourism, gastronomy and other service industries.

6.1. Implications for Regional Tourism Administrators

The main implication for regional tourism administrators is that AI recommendation systems are likely to amplify the destinations that already have the strongest digital visibility, the clearest branding, and, importantly, the most repeated online narratives. Essentially, being a good attraction is no longer enough and being “legible” to AI is now of strategic significance. Places and attractions with strong official pages, consistent “signature” positioning, clear landmark imagery, and repeated mentions across tourism sites are more likely to be recommended.
A second implication is that an organization’s web presence is no longer just for humans or search engines such as Google but also needs to be tailored for AI systems. More than ever, a destination narrative will be influenced both by what the tourism body publishes and by what others repeat about the destination.
Thirdly, AI can reproduce inherited promotional bias and crowd out quieter places in the region covered by a regional tourism body. This raises a destination equity issue: flagship towns may keep winning attention while smaller towns, newer products, and less marketable experiences become harder to discover.
A strategic approach would be to (i) strengthen official destination pages with clear, structured, up-to-date, factual content; (ii) sharpen each destination’s distinctive identity so that the region is not represented only by its biggest anchor destination; (iii) coordinate regional digital narratives across tourism sites, event pages, media, and social channels; (iv) monitor how AI describes a region and where it omits or mis-ranks places; (v) actively support lesser-known destinations with stronger digital packaging, imagery, and authoritative content so they are more “AI-discoverable.”

6.2. Implications for Organizers of Well-Known, but Not Iconic, Christmas Markets

For an organizer of a well-known but not iconic Christmas market, the key challenge is not basic awareness but crossing the threshold from known to “default pick.” AI favors markets with a “clear identity,” “stable home base,” strong flagship branding, and a credible official tourism or event listing.
Being well-known, but not iconic, makes a market vulnerable to being crowded out by hidden filters. The AI model injected its own assumptions about geographic spread, market “vibe,” and trip-planning usefulness. When targeting the “vibe” component, distinctiveness matters more than general quality. The model looked for markets known for something specific, such as a signature food, craft tradition, theme, or setting, rather than simply well-run markets.
AI will reward repeatability and consensus. The model explicitly used consensus and repeated inclusion across sources as a marker of authority. In consequence, the digital footprint must be coordinated, as inconsistent messaging across a market’s website, tourism partners, travel articles, Instagram, Facebook, and Wikipedia can weaken the market’s AI visibility. While the AI model balanced official sites, tourism board pages, major travel outlets, Wikipedia, and informal sources, even when social media was not cited directly, it still influenced the search path and could reinforce the “usual, popular” narrative.
A strategic approach would be to (i) develop a sharp signature with one or two clear attributes people can instantly associate with the market; (ii) maintain a flagship digital page that is authoritative, current, image-rich, and easy for AI systems to interpret; (iii) ensure narrative consistency across channels where the market’s official site, regional tourism pages, media coverage, and social content should reinforce the same positioning and signature; (iv) achieve and maintain third-party repetition by inclusion in credible roundups, travel media, and destination guides.

6.3. Implications for Organizers of Iconic Christmas Markets

The main implication for an organizer of an iconic market that already makes the shortlist is that AI is now helping to lock in its advantage. That advantage, however, survives only as long as the position as one of the default, easiest-to-recommend choices is maintained. The model is rewarding markets that are repeatedly described in “reputation-defining labels,” associated with strong historical continuity, clear symbolic value, and a large-scale or landmark setting. A market’s status depends on keeping its fame machine-readable and constantly corroborated. It follows that the market’s brand is now an asset that must be actively governed by preserving a highly legible, highly repeatable market identity that both humans and AI systems can instantly recognize.
A second implication is that emerging competitors do not need to beat the iconic market on heritage to threaten it. They can attack the position by becoming easier to describe, more photogenic, more digitally coordinated, or more visible across travel roundups and social channels. Since AI relies not only on official sources but also on travel media, blogs, and social media, a rising competitor with sharper online storytelling can close the gap surprisingly fast.
A risk is that AI shortlist status can breed complacency. Because users are shown numbered lists with language like “most iconic” or “can’t-miss,” inclusion creates a perception of objective rank and authority. While that is helpful in the present, it also means any slippage in how a market is described online can alter how future users perceive its importance.
To maintain status and fend off emerging competitors, a solution is to protect a market’s core symbolic advantage by reinforcing the attributes that make it canonical: heritage, tradition, landmark setting, signature rituals, and the sense that missing that market would feel like missing a central part of Christmas-market culture. These are exactly the kinds of cues AI uses as stable reasons for fame. This can be achieved by “owning” the official digital record and standardizing the market’s narrative across the web and social media. Develop and maintain a few phrases and proof points that should be repeated everywhere: what makes the market historic, iconic, distinctive, and unmissable. These same claims need to appear consistently on the market’s own site, partner tourism pages, press materials, travel features, and social bios. Repetition across independent sources is one of the signals that makes a market “canonical” in the eyes of AI.
While merging rivals may win attention by seeming more novel, a market position defense strategy is selective renewal by introducing features that deepen the market’s signature rather than distract from it. As AI favors markets known for “something specific,” the protection of that specificity is paramount. If competitors are gaining traction, those dimensions should be strengthened where iconic markets are hardest to displace: continuity, symbolic meaning, recognizable setting, flagship branding, and bucket-list status.

6.4. Broader Implications

While focusing on Christmas markets as a case example, the study has wider, more generalizable implications. Essentially, organizations now need an “AI visibility” strategy alongside their search, media, and visitor strategy. Generative AI will funnel attention toward already dominant places and narratives.
Consequently, it is necessary to “feed” the sources AI is most likely to absorb. AI triangulates across institutional pages, major travel media, reputable features, and, in practice, also blogs and social platforms. Administrators need to ensure a sustained presence in those ecosystems, not occasional PR hits. Because private sites and social media influence descriptive output heavily, a market brand can be reshaped by repeated informal phrasing. Monitoring how the market being managed or promoted is described across travel blogs, Instagram, Facebook, and listicles will allow correction of weak or generic framing before it becomes the default story.

6.5. Future Work

The study has shown that ChatGPT’s recommendations of German Christmas markets are useful but non-neutral and should be interpreted as “curated,” bias-bearing constructs rather than transparent information retrieval. On a practical side, as the paper is constrained by its narrow scope on the Christmas markets, future work could examine different tourism sectors or service industries to test the generalizability of the findings. Expanding beyond the single case study, longitudinal research could examine how outputs evolve over time with model updates. Further work is also needed to systematically evaluate source quality, bias propagation, and the role of social media in shaping AI recommendations.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Data Availability Statement

The raw data can be accessed at https://doi.org/10.26189/6fa2c5f6-46c1-4e85-ba70-c6e66ac8bab5.

Conflicts of Interest

The author declares no conflict of interest.

References

  1. Almeida-Santana, A., David-Negre, T., & Moreno-Gil, S. (2020). New digital tourism ecosystem: Understanding the relationship between information sources and sharing economy platforms. International Journal of Tourism Cities, 6(2), 335–345. [Google Scholar] [CrossRef] [Scilit]
  2. Altman, S. (2023a). Questions for the record following Sam Altman’s testimony before the U.S. Senate Committee on the Judiciary. Available online: https://openai.com/global-affairs/sam-altman-senate-questions-for-the-record/ (accessed on 22 February 2026).
  3. Altman, S. (2023b). Written testimony before the U.S. Senate. Available online: https://openai.com/global-affairs/testimony-of-sam-altman-before-the-us-senate/ (accessed on 22 February 2026).
  4. Alves, P., Martins, H., Saraiva, P., Carneiro, J., Novais, P., & Marreiros, G. (2023). Group recommender systems for tourism: How does personality predict preferences for attractions, travel motivations, preferences and concerns? User Modeling and User-Adapted Interaction, 33(5), 1141–1210. [Google Scholar] [CrossRef] [Scilit]
  5. Arora, N., Manchanda, P., Aggarwal, A., & Maggo, V. (2025). Tapping generative AI capabilities: A study to examine continued intention to use ChatGPT in the travel planning. Asia Pacific Journal of Tourism Research, 30(10), 1303–1322. [Google Scholar] [CrossRef] [Scilit]
  6. Banerjee, A., Banik, P., & Wörndl, W. (2023). A review on individual and multistakeholder fairness in tourism recommender systems. Frontiers in Big Data, 6, 1168692. [Google Scholar] [CrossRef] [Scilit]
  7. BDSM. (2001). Weihnachtsmärkte als wirtschaftsfaktor für kommunen und tourismus in der bundesrepublik deutschland sowie dessen beitrag zur leistungssteigerung im mittelständischen schaustellergewerbe und markthandel. Bundesverband Deutscher Schausteller und Marktkaufleute e. V. [Google Scholar]
  8. Borràs, J., Moreno, A., & Valls, A. (2014). Intelligent tourism recommender systems: A survey. Expert Systems with Applications, 41(16), 7370–7389. [Google Scholar] [CrossRef] [Scilit]
  9. Burke, R. (2002). Hybrid recommender systems: Survey and experiments. User Modeling and User-Adapted Interaction, 12(4), 331–370. [Google Scholar] [CrossRef] [Scilit]
  10. Calhoun, D. (2025). How we review at time out. Available online: https://www.timeout.com/how-we-review (accessed on 22 February 2026).
  11. Christou, E., Fotiadis, A., & Giannopoulos, A. (2025). Generative AI as a tourism actor: Reconceptualising experience co-creation, destination governance and responsible innovation in the synthetic experience economy. Journal of Tourism, Heritage & Services Marketing (JTHSM), 11(2), 16–41. [Google Scholar]
  12. Condé Nast Traveler. (2026). About Us. learn more about traveler’s editorial policies, testing guidelines, annual awards, and how to contact the team. Available online: https://www.cntraveler.com/info/about-us (accessed on 25 February 2026).
  13. Dai, S., Shao, N., Zhao, H., Yu, W., Si, Z., Xu, C., Sun, Z., Zhang, X., & Xu, J. (2023, September 18–22). Uncovering chatgpt’s capabilities in recommender systems. Proceedings of the 17th ACM Conference on Recommender Systems, Singapore. [Google Scholar]
  14. Deldjoo, Y. (2025). Understanding biases in ChatGPT-based recommender systems: Provider fairness, temporal stability, and recency. ACM Transactions on Recommender Systems, 4(2), 1–35. [Google Scholar] [CrossRef] [Scilit]
  15. Dogru, T., Line, N., Mody, M., Hanks, L., Abbott, J. A., Acikgoz, F., Assaf, A., Bakir, S., Berbekova, A., & Bilgihan, A. (2025). Generative artificial intelligence in the hospitality and tourism industry: Developing a framework for future research. Journal of Hospitality & Tourism Research, 49(2), 235–253. [Google Scholar]
  16. Florido-Benítez, L. (2024). Generative artificial intelligence: A proactive and creative tool to achieve hyper-segmentation and hyper-personalization in the tourism industry. International Journal of Tourism Cities. [Google Scholar] [CrossRef] [Scilit]
  17. Foroughi, B., Vu, H. T. M., & Thaichon, P. (2025). Building trust for sustained generative AI travel adoption. Tourism Review, 81(2), 718–754. [Google Scholar] [CrossRef] [Scilit]
  18. Fujimoto, S., & Takemoto, K. (2023). Revisiting the political biases of ChatGPT. Frontiers in Artificial Intelligence, 6, 1232003. [Google Scholar] [CrossRef] [Scilit]
  19. Gross, N. (2023). What ChatGPT tells us about gender: A cautionary tale about performativity and gender biases in AI. Social Sciences, 12(8), 435. [Google Scholar] [CrossRef] [Scilit]
  20. Guo, P., Liu, G., Xiang, X., & An, R. (2025). From AI to the table: A systematic review of ChatGPT’s potential and performance in meal planning and dietary recommendations. Dietetics, 4(1), 7. [Google Scholar] [CrossRef] [Scilit]
  21. Gursoy, D., Li, Y., & Song, H. (2023). ChatGPT and the hospitality and tourism industry: An overview of current trends and future research directions. Journal of Hospitality Marketing & Management, 32(5), 579–592. [Google Scholar] [CrossRef] [Scilit]
  22. Guzzo, T., Ferri, F., & Grifoni, P. (2022). What factors make online travel reviews credible? The consumers’ credibility perception-CONCEPT model. Societies, 12(2), 50. [Google Scholar] [CrossRef] [Scilit]
  23. Jenny, D. F., Billeter, Y., Sachan, M., Schölkopf, B., & Jin, Z. (2023). Exploring the jungle of bias: Political bias attribution in language models via dependency analysis. arXiv, arXiv:2311.08605. [Google Scholar]
  24. Jung, H., Sharma, A., & Nicolau, J. L. (2026). GenAI in tourism: Who wins, who loses? Tourism Management, 114, 105357. [Google Scholar] [CrossRef] [Scilit]
  25. Jürgens, U. (2014). Weihnachtsmarkt als event. Ga&S, 36(210), 29–37. [Google Scholar]
  26. Kang, S.-E., Kim, M. J., Kim, J. S., & Olya, H. (2026). Can I trust GenAI to plan my next trip? A multi-method approach to optimizing media mix. Journal of Travel Research, 65(2), 335–353. [Google Scholar] [CrossRef] [Scilit]
  27. Khan, S., Azhar, M., & Azeem, S. (2026). Decoding the digital influence: How Trivago’s information ecosystem drives hotel booking intentions? Journal of Hospitality and Tourism Insights, 9(4), 1585–1604. [Google Scholar] [CrossRef] [Scilit]
  28. Knibbs, K. (2024a). Harvard is releasing a massive free AI training dataset funded by OpenAI and Microsoft. Wired Magazine. Available online: https://www.wired.com/story/harvard-ai-training-dataset-openai-microsoft/ (accessed on 22 February 2026).
  29. Knibbs, K. (2024b). Here’s proof you can train an AI model without slurping copyrighted content. Wired Magazine. Available online: https://www.wired.com/story/proof-you-can-train-ai-without-slurping-copyrighted-content/ (accessed on 22 February 2026).
  30. Koçak, B. B. (2026). How AI-generated messages impact consumer behavior in the tourism industry. Journal of Theoretical and Applied Electronic Commerce Research, 21(2), 67. [Google Scholar] [CrossRef] [Scilit]
  31. Kshetri, N., Dwivedi, Y. K., Davenport, T. H., & Panteli, N. (2024). Generative artificial intelligence in marketing: Applications, opportunities, challenges, and research agenda (Vol. 75, p. 102716). Elsevier. [Google Scholar]
  32. Loh, S., Lorenzi, F., Saldaña, R., & Licthnow, D. (2003). A tourism recommender system based on collaboration and text analysis. Information Technology & Tourism, 6(3), 157–165. [Google Scholar] [CrossRef] [Scilit]
  33. Luo, X., Xu, D., Li, Y., & Wan, L. C. (2025). Advancing information search through GenAI: The roles of search type, travel motive and GenAI customization level. International Journal of Contemporary Hospitality Management, 37(5), 1725–1743. [Google Scholar] [CrossRef] [Scilit]
  34. Mainolfi, G., Lo Presti, L., Marino, V., & Filieri, R. (2022). “YOU POST, I TRAVEL.” Bloggers’ credibility, digital engagement, and travelers’ behavioral intention: The mediating role of hedonic and utilitarian motivations. Psychology & Marketing, 39(5), 1022–1034. [Google Scholar]
  35. Marchiori, E., & Cantoni, L. (2011). The online reputation construct: Does it matter for the tourism domain? A literature review on destinations’ online reputation. Information Technology & Tourism, 13(3), 139–159. [Google Scholar] [CrossRef] [Scilit]
  36. Marti, D., Budathoki, A., Ding, Y., Lucas, G., & Nelson, D. (2025). How does acknowledging users’ preferences impact AI’s ability to make conflicting recommendations? International Journal of Human–Computer Interaction, 41(15), 9466–9477. [Google Scholar] [CrossRef] [Scilit]
  37. MedCalc Software. (2026). Comparison of proportions calculator version 22.032. MEDCALC. Available online: https://www.medcalc.org/calc/comparison_of_proportions.php (accessed on 21 February 2026).
  38. Motoki, F., Pinho Neto, V., & Rodrigues, V. (2023). More human than human: Measuring chatgpt political bias. Public Choice, 198, 3–23. [Google Scholar] [CrossRef] [Scilit]
  39. Nakashima, Y., Hirota, Y., Wu, Y., & Garcia, N. (2023). Societal bias in vision-and-language datasets and models. Journal of the Imaging Society of Japan 日本画像学会誌, 62(6), 599–609. [Google Scholar] [CrossRef]
  40. Nguyen, P. M. B., Pham, X. L., & Truong, G. N. T. (2025). The influence of source credibility and inspiration on tourists’ travel planning through travel vlogs. Journal of Travel Research, 64(1), 222–237. [Google Scholar] [CrossRef] [Scilit]
  41. Nicolau, J. L. (2025). When ChatGPT designs your trip: How GenAI adds a cognitive layer to smart tourism. Journal of Smart Tourism, 5(4), 137–151. [Google Scholar] [CrossRef] [Scilit]
  42. O’Connor, P., & Assaker, G. (2024). Examining the role of reputation as a moderator of e-service quality, trust, and e-loyalty in online travel services. Journal of Theoretical and Applied Electronic Commerce Research, 19(4), 3429–3442. [Google Scholar] [CrossRef] [Scilit]
  43. OpenAI. (2022). ChatGPT: Optimizing language models for dialogue. Available online: https://web.archive.org/web/20230131024235/https://openai.com/blog/chatgpt/ (accessed on 31 January 2023).
  44. OpenAI. (2023). GPT-4 Technical Report. arXiv, arXiv:2303.08774. [Google Scholar]
  45. OpenAI. (2024). GPT-4o system card. Available online: https://cdn.openai.com/gpt-4o-system-card.pdf (accessed on 25 February 2026).
  46. OpenAI. (2025a). GPT-5.1 instant and GPT-5.1 thinking system card addendum. Available online: https://cdn.openai.com/pdf/4173ec8d-1229-47db-96de-06d87147e07e/5_1_system_card.pdf (accessed on 25 February 2026).
  47. OpenAI. (2025b). Introducing GPT-5.2. Available online: https://openai.com/index/introducing-gpt-5-2/ (accessed on 18 February 2026).
  48. OpenAI. (2025c). OpenAI GPT-5 system card. arXiv, arXiv:2601.03267. [Google Scholar] [CrossRef] [Scilit]
  49. OpenAI. (2025d). Update to GPT-5 system card: GPT-5.2. Available online: https://cdn.openai.com/pdf/3a4153c8-c748-4b71-8e31-aecbde944f8d/oai_5_2_system-card.pdf (accessed on 25 February 2026).
  50. Rutinowski, J., Franke, S., Endendyk, J., Dormuth, I., & Pauly, M. (2023). The self-perception and political biases of ChatGPT. arXiv, arXiv:2304.07333. [Google Scholar] [CrossRef] [Scilit]
  51. Singh, S. (2026). ChatGPT statistics (2025): DAU & MAU data worldwide. Available online: https://www.demandsage.com/chatgpt-statistics/ (accessed on 3 May 2026).
  52. Smithson, S., Devece, C. A., & Lapiedra, R. (2011). Online visibility as a source of competitive advantage for small-and medium-sized tourism accommodation enterprises. The Service Industries Journal, 31(10), 1573–1587. [Google Scholar] [CrossRef] [Scilit]
  53. Spennemann, D. H. R. (2023). Children of AI: A protocol for managing the born-digital ephemera spawned by generative AI language models. Publications, 11, 45. [Google Scholar] [CrossRef] [Scilit]
  54. Spennemann, D. H. R. (2025a). Commercial Christmas cheer: The reestablishment and proliferation of Christmas markets in post-World War II Germany. Economic History eJournal (SSRN Preprints), 17. [Google Scholar] [CrossRef] [Scilit]
  55. Spennemann, D. H. R. (2025b). Generative artificial intelligence and the future of public knowledge. Knowledge, 5(3), 20. [Google Scholar] [CrossRef] [Scilit]
  56. Spennemann, D. H. R. (2025c). Longer-term effects of COVID-19 on the visitation of German Christmas markets. arXiv. [Google Scholar] [CrossRef] [Scilit]
  57. Spennemann, D. H. R. (2025d). The layered injection model of algorithmic bias as a conceptual framework to understand biases impacting the output of text-to-image models. SSRN. [Google Scholar] [CrossRef] [Scilit]
  58. Spennemann, D. H. R. (2025e). The origins and veracity of references ‘cited’ by generative artificial intelligence applications. Publications, 13, 12. [Google Scholar] [CrossRef] [Scilit]
  59. Spennemann, D. H. R. (2025f). Who is to blame for the bias in visualizations, ChatGPT or Dall-E? AI, 6(5), 92. [Google Scholar] [CrossRef] [Scilit]
  60. Spennemann, D. H. R. (2026a). Examining the usefulness of generative AI as a travel recommender system for German Christmas markets—Supplementary Data. School of Agricultural, Environmental and Veterinary Sciences; Charles Sturt University. [Google Scholar] [CrossRef]
  61. Spennemann, D. H. R. (2026b). Examining the usefulness of generative AI as a travel recommender system for German Christmas markets. Documentation of conversations with ChatGPT5.2. School of Agricultural, Environmental and Veterinary Sciences; Charles Sturt University. [Google Scholar] [CrossRef]
  62. Spennemann, D. H. R. (2026c). Examining the usefulness of generative AI as a travel recommender system for German Christmas markets. Documentation of conversations with ChatGPT5.2 related to data selection and process. School of Agricultural, Environmental and Veterinary Sciences; Charles Sturt University. [Google Scholar] [CrossRef]
  63. Spennemann, D. H. R. (2026d). Examining the usefulness of generative AI as a travel recommender system for German Christmas markets. Documentation of the content of the ChatGPT5.2 reasoning panel. School of Agricultural, Environmental and Veterinary Sciences; Charles Sturt University. [Google Scholar] [CrossRef]
  64. Spennemann, D. H. R. (2026e). Images selected by ChatGPT5.2 to illustrate lists of German Christmas markets. A data set. School of Agricultural, Environmental and Veterinary Sciences; Charles Sturt University. [Google Scholar] [CrossRef]
  65. Spennemann, D. H. R. (2026f). Nature of sources cited in German-language Wikipedia pages on German Christmas markets. School of Agricultural, Environmental and Veterinary Sciences; Charles Sturt University. [Google Scholar] [CrossRef]
  66. Spennemann, D. H. R., & Parker, M. (2021). The changing face of German Christmas Markets: Historic, mercantile, social and experiential dimensions. Heritage, 4(3), 1821–1835. [Google Scholar] [CrossRef] [Scilit]
  67. Spennemann, D. H. R., & Parker, M. (2025). “Endlich wieder Strietzeln”: The restart of German Christmas markets after the COVID-19 pandemic. International Journal of Event and Festival Management, 16(4), 509–529. [Google Scholar] [CrossRef] [Scilit]
  68. Spennemann, D. H. R., Parker, M., & Bond, J. (2025). Marketing Christmas markets: Emotive and commercial domains in promotional texts. International Journal of Tourism Cities, 12(1), 117–134. [Google Scholar] [CrossRef] [Scilit]
  69. Sullivan-Paul, M. (2023). How would ChatGPT vote in a federal election? A study exploring algorithmic political bias in artificial intelligence [Ph. D. thesis, University of Tokyo]. [Google Scholar]
  70. Szymański, P., Lipczyńska, M., & Górska, A. M. (2025). From data to perception: Visualizing bias in artificial intelligence-generated images. European Heart Journal, 46(19), 1781–1782. [Google Scholar] [CrossRef] [Scilit]
  71. Tan, W.-K., & Chang, Y.-C. (2011). Credibility assessment model of travel information sources: An exploratory study on travel blogs. Springer. [Google Scholar]
  72. Thomson, E., Chapman, R., Cooper, G., & Cooper, Z. (2025). Exploring bias in artificial intelligence: Stereotypes and gendered narratives in digital imagery of early childhood educators. Gender and Education, 38(1), 53–72. [Google Scholar] [CrossRef] [Scilit]
  73. Travel, A. (2023). Code of ethics for archaeology travel. Available online: https://archaeology-travel.com/code-of-ethics/ (accessed on 22 February 2026).
  74. Wang, G., Yang, R., Zhang, Y., Wen, X., Liu, C., Liu, E., Tang, M., Xue, L., & Liu, Z. (2025). Evaluating the performance of Large language models in rheumatology for connective tissue diseases: DeepSeek-R1, ChatGPT-4.0, Copilot, and Gemini-2.0. International Journal of Medical Informatics, 207, 106172. [Google Scholar] [CrossRef] [Scilit]
  75. Williams, A. (2025). About amanda & a dangerous business. Available online: https://www.dangerous-business.com/about-2/ (accessed on 22 February 2026).
  76. Wintertraum GmbH. (2024). Die schönsten Weihnachtsmärkte 2024 in Deutschland entdecken. Available online: https://www.deutsche-weihnachtsmaerkte.de/weihnachtsmarkt/ (accessed on 12 December 2024).
  77. Wong, J. W. C., Lai, I. K. W., & Lin, Y. (2025). The perceived reliability and adoption intention towards human-generated content vs. AI-generated content for travel planning: A moderating role of travel persona. Journal of Travel & Tourism Marketing, 42(4), 461–478. [Google Scholar]
  78. Xiang, Z., Pan, B., & Fesenmaier, D. R. (2009, June 21–24). Benchmarking the visibility of websites in Google: Implications for search engine marketing of tourism destinations. TTRA Annual International Conference 2009, Honolulu, HI, USA. [Google Scholar]
  79. Zhu, Z., Kim, S., & Shin, H. H. (2022). Does the website credibility of online travel agencies (OTAs) affect customers’ evaluation and future intentions? Asia Pacific Journal of Tourism Research, 27(8), 842–855. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Journal Statistics
Multiple requests from the same IP address are counted as one view.
“Emergence” and “Dissolution” of Green Innovation Bubbles in Power Industry Chain Enterprises
Previous
Customer Incivility Spillover into Kitchen Staff Deviance and Withdrawal in Multigenerational Workplaces: The Moderating Function of Moral Disengagement
Next