Exploring Shifts in User Behavior Through Longitudinal Data from a Digital Platform for Art and Culture

Abstract

Digital repositories have been an important gateway for the dissemination of information regarding objects of art and cultural heritage throughout the World Wide Web, but the vast number of available artifacts, both historical and modern, makes their discovery by interested users an arduous task. Often deviating from general-purpose search behavior, people searching online for art and culture adjust their habits to address this challenge. In this study, real-world data from the federated search engine and online art and culture repository ArtBoulevard are used to explore this evolution throughout a period of three years. By collecting and analyzing a large amount of user session data, this research aims to investigate user engagement, query formulation, search behavior, and in-platform and outbound engagement in order to outline the longitudinal behavioral patterns of the platform’s user base. Over the period of the analysis, shifts and trends are identified and discussed within the ever-evolving context of behavioral analysis in the field. This process leads to useful insights that are not only indicative of the platform’s limited but global user base, but which can be useful to all stakeholders active in content dissemination and may also be relevant to broader discussions about the changes in the discovery pathways in art and cultural heritage.

1. Introduction

Since the dawn of the digital era, cataloging and preserving items of artistic or cultural importance has been a significant driving force for innovation in the field of digital repositories, through the establishment of standardized metadata and similar best practices [1]. The process of documentation and digital preservation of art and culture is constantly undergoing evolutionary changes with the inclusion of cutting-edge technologies such as VR/AR [2], 3D scanning techniques [3] and artificial intelligence (AI) [4]. This technological progress has led to increased research interest on a global scale, driving interdisciplinary growth within the wider landscape of digital cultural heritage [5]. But the ever-expanding landscape of digital artifacts and their metadata creates hurdles regarding dissemination to members of the global audience and poses new challenges to the user-oriented search and discovery of useful and valid art and culture-related content [6]. In order to address these challenges and design for a better experience, it is paramount that we gain a better understanding of how users from these fields search and engage with digital online repositories [7]. Modern approaches are being used to gain insight into the search process, including monitoring user emotional responses [8]. This understanding often leads to actionable insights that both help with discoverability [9] and show how a more user-focused approach may be employed to ensure accessibility, usability and emotional engagement [10].

Users looking for content directly related to art and cultural heritage have been known to show noteworthy deviations from general Web search behavior in terms of exploration metrics and usage context [7,11]. Art searchers adapt their behavioral patterns with regard to the complexity of their goal and utilize a mix of general-purpose tools and special-purpose repositories [12]. Going beyond search-related behavior, interaction with rich digital collections has also presented some contrast with typical behavioral patterns [13]. Aspects of the user experience (UX) that deal with result prioritization, search refinement and interface personalization have so far been shown to play an important role in art and culture-related interactions [14], while at the same time users appear more vulnerable to UX related problems such as interface clarity [10]. In order to collect more indicators concerning these behavioral patterns, it is important to have access not only to lab studies and self-reported behaviors, but also to actual, real-use data recorded in the wild, where users show their candid behavior.

Since the early days of the Web, the observation of real users has been a cornerstone of behavioral analysis, specifically with regard to the online searching user experience [15]. Towards gaining a better understanding outside a controlled environment, transaction log analysis became a staple methodology for collecting actionable real-world data [16]. Specifically, search log analysis can help detail the information searching tactics of users and lead to improvements in information system design [17]. This approach has been a continuous force behind user-focused research in the field of digital repositories, with multiple works clearly indicating the value of real data and how a setting that encourages natural usage provides higher accuracy than lab experiments [9]. Carefully planned real-world data collection of behavioral data addresses issues of balancing precision, unpredictability and interpretability, while at the same time working within the ethical constraints dictated by respect for privacy and individuality [18]. A very important part of candid data collection is its ability to factor real search behavior, as a process that spans multiple platforms, including both general-purpose popular search systems like Google Search and content providers such as institutional repositories [19]. General-purpose tools often act as an entry point towards the original digital repository, but this behavioral pattern creates challenges for content providers [19]. Looking at each individual platform as part of the wider Web environment is also an important consideration of this article.

In the past few years, the Web analytics field has been facing a new challenge: bot traffic on the Web is rising, motivated by the training needs of AI models [20]. This impact is becoming apparent more and more in cultural repositories, and addressing the traffic regime it brings can be a complicated task [21]. On the other hand, a shift towards AI-based Web search has also manifested in these last few years, with generative models becoming an ever more popular and viable alternative to traditional keyword-based Web search [22]. The use of Large Language Models (LLMs) to enhance cultural and artistic search through repositories such as Europeana is an emerging field of interesting research [23].

Despite the inherent importance of real data in and of itself, they represent essentially nothing more than a snapshot in time. Going one step further to encompass the wider picture, the parameter of time is also paramount. In order to take this into consideration, research in the field of Human–Computer Interaction (HCI) has strived since the early days to explore field-based longitudinal observation as a means to study user behavior [24]. Especially when human behavior is involved, contextual factors, as well as personal development of individual opinions and attitudes, drive behavioral shifts that evolve over time, and their understanding is important in assessing the processes that lead to change [25]. In the fields of art and culture, longitudinal studies have been employed to assess user engagement with cultural platforms while accounting for the dynamic nature of artistic content [26]. Leveraging the importance of usability in established digital art and cultural heritage platforms [11,14] and the distinct needs of people searching online for art and culture [7], alongside the accuracy of real-world collected data [18] and the power of long-term analysis in revealing the complexity of repository-related behavior [27], it becomes apparent that this avenue of research has much to offer. Taking all this into consideration, the present study aims to explore the stability and shifts in user behavior during their interaction with the ArtBoulevard federated search engine and digital art and culture repository platform. By taking advantage of thousands of interaction sessions candidly performed “in-the-wild” from around the globe over the course of the past three years, the study focuses on investigating the following research questions:

RQ1. How did user engagement evolve over the three-year period?

RQ2. To what extent did general query search behavior change over time?

RQ3. How did in-platform result engagement and outbound engagement shift over time?

This research’s ultimate goal is to use its wide array of collected interaction data to identify and assess longitudinal stability and behavioral shifts as they occurred on a live public platform, available freely on the Web, and through that, gain insights that may be useful to stakeholders from the field of digital art and culture repositories, and relevant to ongoing changes in how users search and discover art and culture online. Gaining more candid knowledge about user engagement, query search behavior and in-platform and outbound engagement can help art and culture content providers streamline their online presence in order to better address the needs of visitors and tailor the provided user experience (UX) to their habits. This study highlights the complexities of user search behavior, and its findings can have practical implications for designers and developers in the field of art and culture dissemination. As engagement patterns shift, digital repositories must constantly reassess their UX in order to maintain and increase outreach in an ever-more competitive landscape, especially with the major behavioral shifts brought on by the increased prevalence of AI.

2. Methodology

2.1. Research Design

The research presented in this article was carried out in two distinct phases. In the first phase, the large logs containing information about every user session during the three-year period were programmatically processed, and distinct metrics were calculated for every six-month subperiod that belonged within that timeframe, with each individual year split into two: the first semester from January to the end of June and the second one from July to the end of December. The analytics collection process of the ArtBoulevard platform did not just collect information about query formulation as is customary in similar types of log analysis [28] but instead included a wide array of data regarding content engagement and platform interaction [7]. Interaction information that encompasses the complete input a user provides to the platform is very valuable in the context of digital heritage systems [29]. User behavior in digital repositories can be assessed not only on the basis of fundamental behavioral engagement metrics, such as session duration, but also through discovery engagement metrics based on content interaction [30], such as result views. In the present study, a select few metrics out of the vast array of total actions recorded in the platform’s logs were singled out and used to assess the most important aspects of user interaction.

In the second phase of the research, the aggregated information regarding every semester covered by the study was analyzed through the use of descriptive statistics. The analysis focused on both intensity metrics, numerically measuring the full extent of a type of interaction, and prevalence metrics, recording how often a behavior occurs across sessions. The usage of both types of metrics is a staple in methodologies investigating user behavior through quantitative means in various interdisciplinary fields, including education [31], healthcare [32] and social media analysis [33]. Intensity metrics include counts for interactions or quantitative measurements such as time, while prevalence metrics identify proportionality and rates. For intensity metrics, the study’s analysis focused on using quantiles, a well-established fundamental tool of statistical analysis, as they are much more robust to outliers than using means or variance [34]. The accuracy of various metrics, both intensity and prevalence-based, was investigated by the estimation of confidence intervals in an effort to paint a more informative picture through this widely accepted method in applied statistics [35]. Specifically for proportions, Wilson’s score interval was used for its simplicity and performance even for smaller samples [36].

Overall, the research design focused on using established techniques both for data collection and for analysis, aiming to provide a simple methodological workflow that would shift the focus to the longitudinal findings themselves. Figure 1 presents an overview of the research design process.

2.2. The ArtBoulevard Platform

In order to have a clearer understanding of the presented research, it is essential to discuss a bit about the ArtBoulevard platform, which was the digital tool that allowed data collection. ArtBoulevard, available online at https://artboulevard.org/ (accessed on 26 March 2026), is a Web and mobile application that works as a federated search engine. The platform allows users to perform text-based queries and then collects results from a multitude of online art and culture repositories through their publicly available application programming interfaces (API). These repositories include:

• The Europeana Collections;
• The Harvard Art Museums;
• The Metropolitan Museum of Art;
• The Danish National Gallery of Art (SMK);
• The Artsy Online Art Marketplace;
• The CrossRef Scholarly Texts Archive;
• The Open Library of Books.

The collected results focus on works of art or objects of cultural heritage, but also include academic articles and books relating to such works. Figure 2 presents the homepage of the ArtBoulevard platform.

For each discovered item, the platform generates a specific details page with its own unique URL, a textual description of the work, its available metadata and an image which is often a digital representation of the work itself. Another major part of the interface is a direct link to the original source of the item on the appropriate website of its provider, which appears in a prominent spot near the top of each item’s detail page. The ArtBoulevard platform allows simultaneous search in multiple repositories, enhances content through accessibility features, and continuously increases its informational base in a sustainable manner, as the content provided by the various APIs increases. This way it ensures a useful and usable experience for all people interested in art and culture-related content.

It should be noted here that the platform did not undergo any significant UI/UX changes during the three-year period of this study, so it is safe to evaluate the collected analytics without defining breakpoints where the provided interaction underwent serious alterations. That being said, changes in the system administration environment during that period may still have had ramifications that may in turn have influenced traffic quality and quantity. Such changes included hosting environment software updates that may have allowed increased discoverability, as well as changes in firewall security that may have excluded potential visitors. None of these was on a scale where it would be expected to have a large impact.

2.3. Data Collection

As already hinted in the research design section, the platform collects an extensive array of user interactions dealing not only with query formulation but also with all the interactivity offered by the platform such as mouse and scroll movement, system preferences, result views, usage of the accessibility features like text-to-speech, use of the voice search or advanced search functionality, perusal of the platform’s informational pages or announcements and more. As soon as a user navigates to a webpage of the platform, their session is initiated, and every action is recorded alongside its timestamp in a timeline table that is part of the platform’s integrated relational database [7]. Using this structure, a user session can be recreated accurately to a great extent through the recorded analytics information. All this information is anonymized and held locally and in full accordance with the EU’s General Data Protection Regulation (GDPR). Each session is considered over when a user stops interacting with the platform in any way for longer than 20 min, since this time of inactivity is reasonably enough for any further interaction to be considered a new session.

For the purposes of this longitudinal study, all sessions that started after 1 January 2023 and before 31 December 2025 were explored. From each session, only specific metrics were identified and derived from the platform’s timeline table in order to keep the study’s scope manageable. These are presented in

Table 1. Investigated per-session metrics and their descriptions.

Metric	Description	Type
words	The average number of keywords for all of the queries in a session where a query was performed	Numerical
duration	The duration of a session in seconds	Numerical
total_actions	The total number of interactions recorded in the timeline *	Numerical
queries	The number of queries in sessions where at least one query was performed	Numerical
results	The number of result details viewed in sessions where at least one result was viewed	Numerical
engaged_session	True for sessions that lasted at least 3 s or displayed captured mouse or scroll movement	Boolean
query_session	True for sessions where at least one query was performed	Boolean
result_session	True for sessions where at least one result was viewed	Boolean
source_session	True for sessions where at least one external link to an original source repository was clicked	Boolean
mobile_session	True for sessions where the user’s device was identified as a mobile phone	Boolean

* Recorded interactions include performed queries, viewed results, viewed source links, clicks on related search terms, user profile views, preference changes, use of text-to-speech accessibility features, and query and result bookmarking [7].

alongside a short description and the type of metric.

For the purposes of the result_session metric, a result is considered viewed when the user lands on this specific results details page within the platform, regardless of the pathway they followed to reach that page.

The thought process behind the selection of these particular metrics was to encompass a wide array of interactions that can broadly describe each session and to avoid very specific interface actions that may be of little value for a wider longitudinal analysis. Participation and overall engagement may be inferred through the engaged_session, mobile_session, total_actions and duration metrics, query formulation and repeatability through the words, queries and query_session metrics, in-platform engagement through the results and result_session metrics, and outbound engagement through the source_session.

During the aggregation process, explicit automated systems (crawlers, spiders, bots, etc.) were completely excluded from the collected data through an algorithm based on the reported user agents. This allowed the removal of a very large number of automated visits. The remaining sessions were evaluated on the basis of engagement in order to further address low-engagement sessions that may be the result of bot activity not explicitly declared in their user agent. Sessions with a duration lower than 3 s were considered bounces or bot-oriented behavior, and although reported at the beginning of the results section, they are not part of the engaged sessions on which the rest of the analysis was based. As the vast majority of bot visits consist of a singular request with up to 1 s duration [37], while common real human behavior considers the content for 6–8 s before either abandoning or clicking through [21], the threshold of 3 s was selected to be in the middle ground between these values. It should also be noted that any session duration measurement from the platform requires at least one more interaction with the platform (in the form of an HTTP request) after the initial request, as singular instant events by definition do not have durations.

With the above metrics recorded for every session in a distinct and tailor-made database table, a secondary process aggregated the information per semester. Six full semesters were established between 1 January 2023 and 31 December 2025, as well as 11 rolling semesters with an overlapping three-month window that, even though they contain duplicate information, can be used to make it easier to present the points in time where trends shifted.

2.4. Statistical Analysis

In order to pave the way for the statistical analysis, the numerical type values of Table 1 were aggregated into intensity metrics, while the boolean type values were aggregated into corresponding prevalence metrics. with regard to intensity metrics, the three quartiles Q1, median, and Q3 which indicate the 25th, 50th and 75th percentile were calculated. The interquartile range (Q3-Q1) was used to describe the distance between the Q3 and Q1 values, which represents the middle 50% of the data. Finally, the accuracy of the median value was gauged using confidence intervals (CI). The nonparametric bootstrap method was used to establish the CIs for each median. Through this, multiple bootstrap samples of similar size are formulated, and each sample’s median is calculated. Finally, the P2.5 and P97.5 percentiles of the resulting distribution are used as the 95% CI [38].

With regard to prevalence metrics, the total number of sessions for which the corresponding boolean per-session metric held true was counted. The proportion of these sessions out of the total number of engaged sessions was calculated as a percentage, with the exception of the engaged_session metric, which used the total number of sessions as the denominator. Since these metrics are proportion-based, their CI was calculated on the basis of the Wilson score confidence interval for binomial proportions [36].

Table 2. Investigated parameters for every per-session metric and their descriptions.

Metric	Description	Used for
Q1	Value for which 25% of data points are below	Intensity Metrics
median	Value for which 50% of data points are below	Intensity Metrics
Q3	Value for which 75% of data points are below	Intensity Metrics
IQR	Differences between Q3 and Q1, representing 50% of the data	Intensity Metrics
Median 95% CI Low	Lower bound of the 95% bootstrap percentile confidence interval for the median	Intensity Metrics
Median 95% CI High	Higher bound of the 95% bootstrap percentile confidence interval for the median	Intensity Metrics
N	Total number of sessions that hold true for the specific per-session metric	Prevalence Metrics
p%	Percentage of sessions that hold true for the specific per-session metric out of all engaged sessions	Prevalence Metrics
p% Wilson CI Low	Lower bound of the 95% Wilson score confidence interval	Prevalence Metrics
p% Wilson CI High	Higher bound of the 95% Wilson score confidence interval	Prevalence Metrics

presents all aggregated metrics, which were established per semester based on the per-session metrics presented in Table 1.

It should be noted that, as made clear through the descriptions of Table 1, two numerical metrics, namely the results and queries, are only counted for sessions where at least one instance of each event actually occurred. Thus, when their quantiles are calculated for each semester, they become a conditional intensity metric, which aims to avoid zero-inflation for zero-heavy metrics. The conditional intensity metrics, in tandem with their corresponding proportional metrics, which are derived from the r_session and q_session per-session metrics, help shed a more insightful glimpse with regard to this behavior by combining both prevalence and intensity when true.

All of the individual parameters of Table 2 were calculated for each full semester completely independently of any other semester.

3. Results

Throughout this section, tables and graphs presenting the accumulated data are presented, organized on the basis of traffic and engagement, text query behavior, in-platform engagement, and outbound engagement. The findings form the basis on which the research questions are discussed further in the discussion section that follows. Only a portion of the metrics will be presented in this chapter, while the full metrics are available as a spreadsheet at https://zenodo.org/record/19257101 (accessed on 30 March 2026) alongside the full dataset of the study in the form of mySQL relational database table files.

3.1. Traffic and Engagement

In order to discuss traffic, first, we take a look at the total and engaged sessions and the ratio between them.

Table 3. Metrics describing platform traffic.

Semester	Total Sessions N	Engaged Sessions N	Engaged Sessions p%	Engaged Sessions CI Low	Engaged Sessions CI High
1	4672	1507	32.26%	30.93%	33.61%
2	6587	1238	18.79%	17.87%	19.76%
3	12,558	1249	9.95%	9.43%	10.48%
4	19,300	2020	10.47%	10.04%	10.91%
5	72,877	9386	12.88%	12.64%	13.12%
6	2,709,409	194,056	7.16%	7.13%	7.19%

presents the metrics that are essential in this analysis, and Figure 3 presents graphs of the related data.

Throughout the six semesters that the data collection period covered, a gradual increase in traffic was identified, turning into an acute increase in the last two semesters. The same was true for engaged sessions, whose total number slightly drops in semesters 2 and 3 and then increases. Both these metrics indicate an increase in incoming traffic to the platform, although traffic quality may also be shifting over time. The increase in traffic is a sign that the demand for art and culture content is increasing. As the total number of sessions rose, though, the ratio of engaged sessions to total sessions displayed a downward trend. This indicates that the increase in traffic carried with it a higher percentage of low-engagement sessions. As discussed in the following section, this type of reduced engagement trend carries connotations about wider shifts in the art search environment.

Although the bot detection system of the platform remained strict and consistent throughout, it is hard to discern whether all of this traffic was human-related or whether some was the result of automated systems trying to appear as human traffic, especially considering the rapid emergence of AI bots [20,21]. While the noticed change in traffic is consistent with this phenomenon, the study itself cannot infer the origin of the traffic spike with its available data, and changes in the hosting environment mentioned in Section 2 seem unlikely to have had such a major impact. After this first introductory metric, the focus will lie solely in examining engaged sessions, as bounces provide little value in terms of user behavior analysis.

Moving on to more explicit engagement metrics, we take a look at duration, total actions and mobile device use.

Table 4. Metrics describing engagement.

Semester	Duration Median (Seconds)	Duration IQR (Seconds)	Mobile p%	Total Actions Median	Total Actions IQR	Total Actions Q3
1	7	41	23.76%	2	4	5
2	6	26	24.31%	2	3	4
3	14	197	13.53%	2	6	7
4	3	17	14.60%	2	2	3
5	89	775.75	30.43%	2	1	2
6	4	5	14.32%	1	0	1

includes the corresponding metrics, and Figure 4 displays graphs of engagement-related data.

The median duration of sessions shows little fluctuation throughout the data collection period, with the exception of a rather large spike in semester 5. Mobile usage also shows a fluctuating but consistent evolution, especially taking into account the confidence intervals, which also spike in semester 5. This may reflect a short-term change in traffic composition during the 5th semester, in which increased mobile usage may have influenced the duration metric.

Total actions through their median show stability in user engagement with all aspects of the platform. Through the evolution of the IQR distance and the Q3 values of total actions, we can identify a trend towards less engagement per engaged session. In tandem, engaged sessions (Figure 3b) and total actions (Figure 4c) indicate a drop both in the prevalence of engaged sessions and in the intensity of engagement within each one. This downward trend emerged before the large spike in traffic and continued through it, suggesting that it may not be explained solely by the later traffic regime change. These trends may be related to attention span shifts and wider engagement volatility on the Web and are discussed later.

3.2. Querying Behavior

In order to evaluate how the keyword querying behavior of users evolved, we take a look at the prevalence metric of sessions with at least one query, the conditional intensity metric of the number of queries per session with at least one query, and the keywords per query metric.

Table 5. Metrics describing query-related behavior.

Semester	Query Sessions N	Query Sessions p%	Keywords Median	Keywords IQR	Queries Median	Queries Q3	Queries IQR
1	225	14.93%	1.38	1	2	3	2
2	155	12.52%	1	1	1	2	1
3	101	8.09%	1.835	2	1	2	1
4	146	7.23%	2	2	1	2	1
5	248	2.64%	2	2	1	1	0
6	1127	0.58%	3	2	1	1	0

presents these metrics, and Figure 5 presents graphs of query behavior-related data.

As the semesters go by, the percentage of sessions that include at least one query decreases. This trend is apparent even before the traffic regime change in semester 6. Moreover, the number of keywords per query shows an upward trend, further supported by the metric’s IQR, which may be related to a shift towards more natural phrasing in queries.

In terms of query-related intensity, the median displays stability in the last semesters, with single query sessions being the norm. Taking a closer look at more engaged sessions through their Q3, we identify a downward trend, which is in line with the similarly decreasing total actions Q3 (Figure 4c).

3.3. In-Platform Engagement

In order to draw insights concerning in-platform engagement, we focus on the combination of the prevalence metric of sessions with at least one result viewed and the conditional intensity metric of the number of results for sessions with at least one result viewed. It is important to note that viewing a result in this context indicates navigating to that specific item’s details page within the ArtBoulevard platform.

Table 6. Metrics describing in-platform engagement with results.

Semester	Result Sessions N	Result Sessions p%	Results Median	Results Q3	Results IQR
1	677	44.92%	1	2	1
2	675	54.52%	1	2	1
3	932	74.62%	1	4	3
4	1566	77.52%	1	1	0
5	6926	73.79%	1	2	1
6	176,476	90.94%	1	1	0

presents the corresponding metrics, and Figure 6 presents graphs of the result-related data.

The total number of sessions with at least one result viewed shows an increase in engagement in absolute numbers. The same is true for the percentage of sessions with at least one result viewed out of the total engaged sessions, indicating an upward trend. This increase in prevalence for sessions with at least one result viewed (Figure 6b) comes in contrast with the decrease in prevalence for sessions with at least one query (Figure 4a), indicating that despite the waning popularity of keyword search within the platform, in-platform result engagement is on the rise. This trend is visible both before and after the traffic regime change in semester 6.

Looking at the median number of results for each session with at least one result viewed, we see a stable behavior at exactly one result. Coupled with the corresponding IQR, we see that sessions tend to be limited to one result, with a small fluctuation without disrupting this stability. In order to gain a better glimpse at how this intensity works in more engaged sections, we examine the Q3 quantile, which, similarly to IQR, displays some fluctuation between semesters and also indicates a light downward trend. Over time, more results are discovered, more sessions include at least one viewed result, the median remains at one result per session, and higher-quantile values do not suggest broader within-session engagement. Single result sessions are a very important part of the platform’s engagement and may indicate known-item views or discovery pathways outside the platform, driving traffic towards it.

3.4. Outbound Engagement

As mentioned, an important element of the ArtBoulevard platform is the connection between each item and its original source in the various source repositories. In order to assess how this outbound engagement evolves, we take a look at the sessions where at least one external link to an original source repository was clicked. In addition, the ratio between sessions with a result (in-platform engagement) and sessions with an external link clicked (outbound engagement) was calculated and presented as the exit-through percentage.

Table 7. Metrics describing outbound engagement behavior.

Semester	Source Sessions N	Source Sessions p%	Source Sessions CI Low	Source Sessions CI High	Exit Through
1	115	7.63%	6.40%	9.08%	16.99%
2	76	6.14%	4.93%	7.62%	11.26%
3	103	8.25%	6.85%	9.90%	11.05%
4	218	10.79%	9.51%	12.22%	13.92%
5	296	3.15%	2.82%	3.53%	4.27%
6	801	0.41%	0.39%	0.44%	0.45%

presents the corresponding metrics, and Figure 7 presents graphs of outbound engagement-related data.

As seen in Figure 7, the total number of sessions with at least one outbound source click displays a clear upward trend, especially in the last semesters. On the other hand, the percentage of sessions with such an event shows a downward trend that is reinforced by the traffic regime change in semester 6. The exit-through graph is consistent with this notion, showing fewer and fewer result views leading to a source link click. The increased traffic led to more users navigating to the original repositories over the course of time in absolute numbers, but the exit rate itself is diminishing in a way that corresponds with similar trends noted above. This may be an indicator that, despite the platform’s federated search functionality, its real-world use becomes ever more similar to that of an end-point content provider.

In order to more clearly depict the engagement trends that represent the in-platform and outbound engagement, eleven overlapping semesters with a rolling window of three months were used to create the graphs of Figure 8.

These graphs, which include overlapping sessions, are not meant for comparison but as a more descriptive indicator of when shifts occur. The query, results, and source session graphs reinforce the contrast between increasing result-view prevalence and decreasing querying and outbound-click prevalence. This contrasting finding is discussed in detail with regard to its ramifications concerning the way users interact with the platform in the discussion section below.

In order to compare the prevalence of behavioral shifts and identify if these changes were significant, a Chi-Square [39] test was applied to the query sessions, result sessions and source sessions as they appear in Table 5, Table 6 and Table 7, respectively. A first comparison was made between semesters 1 and 5, and then another one between semesters 1 and 6.

Table 8. Chi-Square test results.

Metric	Semester 1 Proportion	Semester 5 Proportion	Semester 6 Proportion	Chi-Square from 1 to 5	Chi-Square from 1 to 6
Query Sessions	14.93%	2.64%	0.58%	469.09	4464.04
Result Sessions	44.92%	73.79%	90.94%	511.92	3707.93
Source Sessions	7.63%	3.15%	0.41%	70.47	1655.84

presents the results of this comparison.

The Chi-Square test confirms the significance of the downward trends in query and source sessions, as well as the significance of the upward trend of the result sessions. These significant trends appear even between the first five semesters, where the traffic regime change hadn’t occurred. The rather extreme values of the test between semesters 1 and 6 indicate the profound impact of the traffic spike.

4. Discussion

4.1. RQ1: How Did User Engagement Evolve over the Three-Year Period?

Both traffic and basic engagement (Figure 3) displayed an upward trend culminating in an extreme spike in traffic during the sixth semester. As the platform established itself in the art and culture landscape and may have become more discoverable over time, with domain age being a significant factor for SEO [40], its total traffic gradually increased throughout semesters 1–5 until it skyrocketed in semester 6. On one hand, the initial gradual increase in traffic of the first few semesters is compatible with a general trend of increased traffic in digital art and culture repositories in the post-COVID-19 era [41]. On the other hand, the spike in the last semester indicates a significant shift in the platform’s traffic regime. The new increased traffic displayed different qualities than the traffic up to that point, which may reflect a change in traffic composition, the causes of which cannot be determined from the present dataset alone. Some prior literature has discussed the possibility that AI-mediated discovery may cause fluctuation in such engagement metrics [42], while at the same time, advanced bots that are used to train LLMs are increasingly causing disruptions in art and culture repository traffic [21]. A combination of the above might be one possible explanation for this shift.

As the overall number of sessions increased (Figure 3a), so did the proportion of bounce sessions, leading to a significantly reduced percentage of engaged sessions (Figure 3b). This behavior is consistent both before and after the traffic regime change. One possible contributing factor is broader difficulty in sustaining digital attention [43]. The median of total actions per session showed significant stability at around two actions per session (Figure 4c). Low page views per session and low session depth have been noted in recent relevant studies in the context of cultural heritage content [44]. Looking deeper at the Q3 for total actions (Figure 4c), we can see that the trend of reduced engagement continues to emerge in long-tail quantiles. As the Web overall consolidates to a small number of larger platforms [45], this can have an effect on smaller-scale providers like the ArtBoulevard platform. A secondary peak in engagement was noticed in the percentage of engaged sessions (Figure 3b) and session duration (Figure 4a) at the 5th semester, corresponding with a similar peak in mobile phone usage (Figure 4b), which is consistent with earlier observations on the platform linking mobile use with higher engagement [7].

Overall, a steady initial traffic increase and a traffic spike in the sixth semester were noted, accompanied by stability in median engagement metrics, whose higher quartiles, though, hint towards per-session engagement decrease. The signs that traffic is strong and getting stronger give small and medium digital art and culture content providers hope that they can reinforce their position in the wider ecosystem, but the changes in engagement metrics show that the role of such repositories should shift to better accommodate these emerging behaviors. The numbers speak greatly to the importance of the industry adjusting to this new reality.

4.2. RQ2: To What Extent Did General Query Search Behavior Change over Time?

Specifically monitoring the evolution of query search behavior, an equivalent decrease in the prevalence of sessions involving keyword-based search is apparent (Figure 5a). While the traffic increase discussed in RQ1 brings with it an increase in the absolute value of sessions with a query, especially in the last semesters (Table 5), the percentage of sessions with a query out of all engaged sessions moves downward (Figure 5a). The shift away from traditional text search has been documented before as a result of the rise in dominance of Generative Artificial Intelligence (GenAI) tools [22], with discovery through these tools becoming a replacement pathway for users to reach their intended end result [46], so this connection should be tested in future multi-platform or referral-aware studies.

The number of keywords per query (Figure 5b) follows an upward trend, with a steady increase even before the traffic regime change in semester 6. The increase in median keywords, as well as similar increases in the higher quantiles of the same metric, is consistent with a tendency toward longer, potentially more natural-language-style queries [47,48]. In terms of the number of queries in each session (Figure 5c), a stable finding of one query per session is noticed, with Q3 hinting towards a decrease.

Overall, the decline of sessions with queries is in line with the bibliography that suggests users may be moving away from traditional keyword-based queries [22,46]. Even when they are still engaging with the platform in that manner, they gravitate towards longer queries. Since keyword-based querying has been a staple of online art and culture UX, these findings pose the question of whether that central role of querying should be reassessed, both specifically in the ArtBoulevard platform, but also on the wider art and culture content ecosystem.

4.3. RQ3: How Did In-Platform Result Engagement and Outbound Engagement Shift over Time?

Sessions with at least one result viewed are the only metric that not only continuously grows in absolute numbers (Figure 6a) alongside the traffic increase but also grows in terms of prevalence among engaged sessions (Figure 6b). This is an interesting finding that underlines that, despite the general trend of fewer actions per session observed in RQ1 and RQ2, this trend does not continue when it comes to sessions with at least one result viewed. In the wider search environment, the rise of GenAI tools has led to a paradigm where users receive highly relevant, summarized responses that point directly to sources from such tools, reducing the classical query-to-SERP cycle [49]. This wider phenomenon, alongside other external discoverability factors, may be a possible contributor to the increased prevalence of sessions with at least one result viewed within the ArtBoulevard repository, in contrast to its other reduced engagement metrics.

Through assessing the intensity metrics of sessions with at least one result, we see that, in a way similar to the queries, the stability of the one-result-per-session paradigm persists in the median, although in the Q3 quantile, the downward trend is less pronounced and with more fluctuations (Figure 6c) than the similar downward trend regarding the Q3 of queries per session (Figure 5c). The one-result-per-session model is also compatible with potential traffic arriving at the platform through external discovery pathways, although these pathways cannot be suggested through the presented data.

With regard to outbound engagement, the absolute number of clicks to external links towards the original repositories increases (Figure 7a) alongside general traffic, but the prevalence of sessions with at least one source viewed (Figure 7b) and the exit through ratio (Figure 7c) both show a downward trend. This behavior closely corresponds to similar behavior regarding queries and total actions discussed in RQ1 and RQ2. In an era where search behavior is undergoing rapid changes [49,50], users often do not click through to the original sources due to often misplaced trust in the tools themselves [50,51], and this may be one possible explanation for the observed decrease in exit-through ratio (Figure 7c).

Overall, user engagement with in-platform results is increasing, but this increased result-oriented traffic is driving less and less traffic proportionally to the original sources. Both of these shifts are in line with broader changes in search behavior as presented in the recent bibliography [49,50,51]. It is important to note that the decreasing prevalence of sessions with a query (Figure 5a and Figure 8a), the decreasing prevalence of sessions with at least one external link clicked (Figure 7b and Figure 8c), and the increasing prevalence of sessions with at least one result viewed (Figure 6b and Figure 8b) are visible before the sixth semester and the traffic regime change.

Taken together, the findings of the present longitudinal study of user behavior are compatible with wider changes in online search and discovery, including fluctuations in engagement due to the uncertainties introduced by AI traffic in the digital repository environment [42], as well as changes to user behavior associated with the growing use of modern exploratory search tools [22,49,51]. Although this study is limited to a single platform, its candid real-world data and the insights derived from them aim to contribute to the ongoing effort to understand how art and culture content dissemination is evolving during this transitional period. This understanding is quintessential in providing a revamped user experience that focuses on the searchers’ new emerging habits and the traffic quality shifts in the new era. Failure to adapt may mean decreased relevance in an ever-evolving field, which can have dire ramifications with regard to the outreach of important art and cultural content.

5. Conclusions

The presented research is a longitudinal data analysis based on real-world usage data of the ArtBoulevard federated search engine and digital art and culture repository over the period of three years, between 1 January 2023 and 31 December 2025. Through calculating and aggregating intensity and prevalence metrics dealing with traffic and engagement, query search behavior, in-platform result engagement, and outbound engagement, the study has presented a series of descriptive findings that lead to useful insights about the longitudinal behavioral patterns on this live and freely available art and culture platform. The combination of longitudinal observation with a wide variety of interaction metrics beyond simple query logs is a key contribution of this study. The provided insights, although not indicative of the wider population of art searchers, are suggestive of wider shifts in the search pathways in art and culture-related content discovery.

Overall traffic to the platform saw a steady and later acute increase, leading to similar increases in all useful interaction events in absolute numbers (total interactions, results viewed, queries formulated, external links clicked). Specifically, result-related metrics showed an increase in prevalence among all engaged sessions, while query and source-related metrics dealing with specific search pathways showed a decrease in prevalence, and all these shifts were visible both before the traffic spike (S1–S5) and when taking it into consideration (S1–S6). This differentiation in how different indicators move is important because it points to more complex behavioral changes than simple disengagement. The trends that showed significant resilience despite the acute changes in the platform’s traffic regime may be relevant to broader changes in online art and culture content discovery. Moreover, the identified patterns can have practical implications for UI design in small and medium-scale digital repositories, especially in how they present result details, implement keyword-based query search, and facilitate outbound engagement.

The study is limited by the fact that it investigates the behavior of its own userbase, which may not be representative of the wider art and culture searcher population. Additionally, its descriptive single-platform nature means that it cannot identify external referral pathways. The fact that the platform is a federated search engine drawing from multiple repositories with ever-changing content makes it hard to consider things such as repository size, description depth, metadata richness, and digital representation quality in its user behavior analysis. In order to keep the scope of this work manageable, many different aspects of user interaction were omitted from the final monitored metrics. Wider research, including users from other platforms, may lead to more concrete findings, while a deeper dive into the collected longitudinal dataset can also be interesting. Moving towards the future, a deeper analysis of the dataset, not purely on descriptive statistics, but using machine learning techniques, will be employed to investigate the shift in user archetypal behaviors. Moreover, research on the interrelation between GenAI tools and external source repositories or intermediaries like ArtBoulevard can lead to a better understanding of the changes in how users search for art and culture-related content online.

As the landscape of the Web in general, and more specifically the artistic and cultural digital footprint, is shaped by the emerging technologies of artificial intelligence and the important shifts in user experience design, real-world quantitative user-behavior data can help clarify how discovery patterns are changing during this transitional period. Understanding how people engage with the process of art search, first for specific platforms, and then as a platform-agnostic process, may lead to the development of better tools that will help streamline the provided user experience, bringing the needs of art searchers to the forefront.