Mapping 3D Digital Heritage at Scale: A ChatGPT-Assisted Analysis of Sketchfab’s “Cultural Heritage & History” Models

Abstract

This paper evaluates the platform-mediated importance and impact of 3D cultural heritage models stored on Sketchfab by analyzing user engagement and retention metrics (views, likes, and comments), and provides a comparative assessment across other major 3D platforms. Our primary goal is to understand how cultural heritage content performs in terms of reach, engagement, and reuse conditions, and how platform design and taxonomies shape what becomes visible and measurable. We map Sketchfab’s Cultural Heritage & History ecosystem through a reproducible, API-driven workflow built on public metadata for over 1.37 million models (views, likes, comments, tags, and licences). The results depict a domain in rapid expansion between 2018 and 2025, while also revealing a strongly unequal attention economy: most models receive limited interaction, whereas a small minority concentrates visibility and engagement. The category Cultural Heritage & History shows high endorsement relative to reach, consistent with “high-value” engagement once content is discovered. Methodologically, large-scale harvesting required automation to manage cursor pagination, intermittent failures, and rate limits (e.g., HTTP 429). In this context, ChatGPT provided essential support by assisting the design and refinement of the extraction and counting algorithm, replacing what would otherwise have required extensive manual counting and verification at a scale that could plausibly take months.

1. Introduction

In recent years, 3D modeling has assumed a central role in the documentation, conservation, and enhancement of cultural heritage [1,2]. Digital photogrammetry and laser scanning techniques now allow for the accurate acquisition of the geometry and texture of artifacts, archaeological sites, and works of art, offering new possibilities for analysis, virtual restoration, and immersive experience [3]. Recent work has also highlighted how web-based open frameworks for interactive 3D artifact visualization can substantially improve the accessibility and public reach of cultural heritage resources, bridging high-fidelity digitization workflows with online dissemination [4]. At the same time, web platforms, like Sketchfab, have revolutionized the dissemination and sharing of 3D models, enabling museums, universities, and researchers to publish interactive models that can be viewed directly from a browser, without the need for dedicated software [5]. Sketchfab has established itself as one of the leading digital infrastructures for the management and dissemination of three-dimensional content, thanks to its intuitive interface, support for various 3D formats (e.g., OBJ, glTF, GLB, USDZ), and the ability to integrate metadata, licenses, and descriptions that facilitate the scientific traceability of content [6]. One of the most significant aspects of the platform is its ability to promote open access and the sharing of 3D models for educational and research purposes. According to Papadopoulos, Gillikin-Schoueri, and Schreibman [7], Sketchfab has not only welcomed but also fostered the growth of 3D content related to cultural heritage, reaching over 100,000 cultural models in 2019, nearly 20,000 of which were distributed under Creative Commons licenses. This openness has enabled the development of new forms of public participation and citizen heritage, where the community actively contributes to the digital documentation of assets. Furthermore, as noted by Khomenko et al., 2025 [8], Sketchfab stands out for its compatibility with web applications and augmented reality, offering a versatile infrastructure for interactive visualization and digital museum education. The platform provides API tools that enable 3D models to be integrated into customized web applications, thus promoting interoperability between digital archives and research environments. In this context, Sketchfab emerges as a global digital laboratory for the understanding, preservation, and dissemination of cultural heritage, contributing to the democratization of access and the construction of a shared cultural memory in three-dimensional format. To ensure consistent classification and structured accessibility of three-dimensional content, Sketchfab employs a comprehensive system of thematic categories. These categories not only organize the vast range of models available on the platform but also mirror the interdisciplinary and creative diversity of contemporary digital production, making Sketchfab a meaningful observatory for analyzing trends in 3D modelling and digital culture. The dataset encompasses a wide range of categories that reflect the thematic and functional diversity of the models available on Sketchfab. Among the most prominent are Animals & Pets and Architecture, covering buildings, urban infrastructures, and both interior and exterior environments. The Art & Abstract category is devoted to artistic creations, digital sculptures, and conceptual compositions, while Cars & Vehicles gathers models of land, air, and maritime transportation means. Further categories include Characters & Creatures, featuring human figures, mythical beings, and imaginary entities, and Cultural Heritage & History, focused on the documentation and preservation of historical and archaeological artifacts. Electronics & Gadgets comprises digital devices and electronic components, whereas Fashion & Style addresses the design of clothing, accessories, and other fashion-related objects. Other thematic groups are Food & Drink, containing culinary items and tools; Furniture & Home, oriented toward interior design and domestic furnishings; and Music, which includes musical instruments and audio equipment. Natural environments are represented through Nature & Plants. News & Politics collects models inspired by current events and socio-political themes; this latter category typically aggregates 3D models produced for political communications or the journalistic/educational visualization of civic affairs (e.g., models of politicians or public figures, campaign materials and party symbols/logos, flags and emblems, protest/signage assets, and other objects or scenes associated with contemporary political events and public debate). Additionally, the dataset features People, with 3D representations of human figures and portraits; Places & Travel, which covers landscapes and touristic reconstructions; Science & Technology, oriented toward educational and research purposes; Sports & Fitness, focused on athletic activities and related equipment; and Weapons & Military, encompassing arms, vehicles, and technologies connected to defence and warfare. Each uploaded model may be associated with a maximum of two categories, ensuring consistency and precision in semantic search and data retrieval. This classification framework enables Sketchfab to maintain a dynamic balance between thematic breadth and descriptive rigor, facilitating both qualitative exploration and quantitative analysis of 3D cultural content. Among the thematic domains represented on Sketchfab, the Cultural Heritage & History category holds particular significance for the purposes of this research. This section of the platform gathers three-dimensional models related to the documentation, preservation, and dissemination of historical and cultural artifacts, encompassing a wide range of materials such as archaeological finds, architectural heritage, museum collections, and intangible cultural assets. The category serves as a digital crossroads where cultural institutions, researchers, and independent creators converge to share models that bridge the gap between tangible heritage and its virtual representation.

Throughout the manuscript, we use the following operational definitions. Reach refers to exposure to the content (primarily view counts and high-view thresholds). Engagement refers to user interactions that indicate approval or discussion (likes and comments, including high-like thresholds). Impact refers to the combined ability to generate reach and engagement, summarized by statistical indices. Finally, reuse conditions refer to licensing and rights metadata that restrict or permit downstream reuse.

1.1. Literature Review: 3D Digitization Pipelines, Web Platforms, and Reuse Challenges in Cultural Heritage

Three-dimensional digitization has matured into a standard component of cultural-heritage research, conservation and outreach. Early and ongoing methodological work establishes photogrammetry (image-based reconstruction) and laser scanning as complementary acquisition methods that make it possible to capture accurate geometry and appearance of artifacts, structures, and sites at multiple scales [9]. Practical workflow papers emphasize reproducible pipelines from image acquisition through SfM/MVS reconstruction, mesh cleaning, texturing, and web optimization, giving cultural heritage practitioners accessible, repeatable protocols for producing high-quality shareable assets [10,11].

A second strand of the literature examines the role of web viewers and repositories in dissemination and reuse. Over the last decade, Sketchfab has become a de facto public venue for embedding interactive 3D models in websites, teaching resources, and museum pages; scholarship and practice papers repeatedly note Sketchfab’s wide adoption by GLAM institutions and its affordances for embeddability, social sharing, licensing controls, and the lightweight inspection of textured models [12,13]. Institutional case studies, for example, projects by the British Museum, university archaeology groups, and national research archives, show how Sketchfab is used to make meshes available to the public, to support teaching, and to increase the visibility of otherwise inaccessible collections. At the same time, empirical analyses of museum-published models reveal heterogeneity in optimization, metadata quality, and paradata (i.e., capture provenance, processing steps), raising concerns about the reusability and scientific reproducibility of models published for outreach rather than research [14]. Recent systematic surveys and reviews of 3D hosting and repositories situate platforms, such as Sketchfab, within a broader ecosystem of institutional archives, research repositories, and commercial services. Champion and Rahaman’s comparative survey synthesizes features across institutional and commercial hosts, highlighting that commercial viewers (including Sketchfab) excel at quick publication, embeddability, and community reach, but tend to vary in metadata richness and preservation guarantees compared to curated institutional archives [6]. Larger reviews also foreground standardization problems, inconsistent metadata schemas, the lack of paradata, and variable licensing practices, all of which complicate large-scale statistical or comparative work on hosted 3D models [15].

Beyond single-project case studies and platform surveys, there is a growing body of work creating large research collections and corpora drawn from Sketchfab to enable quantitative research into 3D models. The Sketchfab 3D Creative Commons Collection (S3D3C) [3] is one such example: it systematically harvested tens of thousands of CC-licensed models from Sketchfab to support benchmarking and analysis of modern textured 3D assets. This work illustrates two important points: first, that Sketchfab is a uniquely large source of heterogeneous textured models (geometry + materials + sometimes animations); second, that harvesting the platform can enable machine learning and statistical studies, but only when licensing, selection criteria and metadata extraction are handled carefully.

Analyses focused specifically on the cultural-heritage subset of hosted models point to mixed strengths. On the positive side, open sharing on Sketchfab increases visibility for collections, supports pedagogy and public engagement, and can accelerate collaborative investigation [12,16]. On the negative side, empirical audits find inconsistent assignment of categories, frequent absence of rich provenance, and variable adherence to licences and download options, factors that complicate reuse for scientific comparison, longitudinal studies, and archiving [14,15]. These findings motivate methodological caution: when researchers harvest or analyze Sketchfab metadata, they often need to combine automated scraping/API pulls with careful manual validation and treat counts or popularity metrics as approximate rather than definitive.

Metadata, licensing, and sustainability are recurrent themes. Papers and community reports stress the importance of standardized metadata and paradata to ensure long-term research value, reproducibility and ethical reuse. Because many projects initially use Sketchfab primarily for outreach, metadata completeness is frequently sacrificed for speed of publication; consequently, several authors call for community standards and for infrastructure (institutional or consortial) to ensure that 3D scholarly outputs remain discoverable and reusable beyond the lifecycle of any single commercial platform.

Finally, studies in the pedagogical and outreach literature show that Sketchfab is widely used as a hands-on tool in teaching and public education, enabling the embedding of manipulable 3D objects in course pages and open educational resources [13]. At the same time, research into the limits of web viewers for analytical tasks suggests that while Sketchfab is excellent for visualization and lightweight inspection, it is less suited as a primary archive for research-grade 3D data unless accompanied by fully downloadable datasets and rigorous provenance.

1.2. Aim of the Paper

The growing availability of interactive 3D models is an important resource for cultural heritage, enabling the remote visualization and digital analysis of artifacts. Accordingly, this study provides a statistical characterization of Sketchfab’s Cultural Heritage & History ecosystem from complementary perspectives: (i) availability and growth (annual model counts) and its relative weight compared with the platform overall and other categories; (ii) user interaction and engagement (views, likes, comments, and derived normalized indicators); and (iii) reuse availability, as proxied by licensing distributions and the prevalence of “Unknown” licences. Therefore, the aims of this study are:

• to quantify the availability and temporal evolution of cultural heritage 3D models on Sketchfab, including their share relative to the broader platform and cross-category baselines;
• to examine popularity and interaction metrics (views, likes, comments) and their distribution;
• to identify the main contributors and publication trends shaping the corpus;
• to assess reuse conditions through licence statistics and to provide a reproducible, API-driven methodological pipeline that can support future large-scale, platform-based studies of 3D cultural heritage content such as the longitudinal monitoring of publication and engagement trends, cross-category and cross-platform comparisons, and rights/metadata audits for reuse and interoperability.

2. Methodology

2.1. Methodological Framework

This study adopts a systematic, data-driven methodological framework to analyze the structure and audience dynamics of 3D content within Sketchfab, with a particular focus on the Cultural Heritage & History category. The methodological design was developed to ensure comprehensive data coverage, analytical transparency, and reproducibility. Following established practices in digital content analytics, the approach integrates automated data acquisition through the platform’s public API, structured metadata extraction, and the quantitative assessment of engagement metrics. To enable consistent cross-category comparisons, all data were normalized and aggregated within a composite indicator, the normalized mean index (NMI), which quantifies the overall capacity of each category to generate user interaction. This multi-stage framework allows for a rigorous evaluation of patterns in content creation, licensing, and audience response, situating the analysis within a broader empirical investigation of digital heritage dissemination and platform-based participation. The pipeline of the methodological approach is shown in Figure 1.

To investigate the characteristics and patterns of 3D content on Sketchfab, a systematic and multi-stage methodological approach was employed. The study design was structured to ensure comprehensive data coverage, maintain high data integrity, and enable robust quantitative analysis. First, relevant metadata for a wide array of 3D models were collected through the platform’s public API. Subsequently, a dedicated extraction pipeline was implemented to automate data retrieval and manage potential inconsistencies. Finally, the assembled dataset underwent rigorous statistical analysis and visualization to uncover trends in both model popularity and user engagement. The following sections detail each of these phases, highlighting the procedures and considerations undertaken to ensure methodological rigor and reproducibility.

2.2. Data Collection

The present study leveraged the public API of Sketchfab to systematically collect a comprehensive dataset of 3D models. Access to the API was facilitated through a personal API token, ensuring the authenticated and secure retrieval of metadata. For each model, a broad spectrum of descriptive and engagement-oriented attributes was extracted, including a unique identifier, the model’s title, the author’s username and profile link, and the date of publication. Additionally, engagement metrics based on user interaction (i.e., number of views, likes and comments) were recorded, together with associated tags, license information, and the direct URL to the model. This multidimensional data collection strategy was designed to capture both intrinsic content characteristics and patterns of audience interaction, thereby enabling a nuanced analysis of the platform’s ecosystem.

2.3. Data Extraction Procedure

We developed a Python script to systematically collect metadata on 3D models from the Sketchfab REST API (endpoint/v3/models). ChatGPT v.5.2 (OpenAI) was used as a support tool during development: it assisted in drafting the initial code structure, suggesting implementation details for cursor-based pagination and retry/backoff logic, and refining the documentation of the data-collection workflow. All final implementation choices, parameter settings, and code changes were reviewed and validated by the authors. The script queries the cultural-heritage-history category and restricts results to models published within a predefined range of years (here: 2018–2025). For each year, the script sends authenticated HTTP GET requests with parameters specifying the category, the publication date interval (published_since, published_until), and sorting by publication date. Results are retrieved page by page using the cursor-based pagination provided by the API. To ensure robustness, the script implements error handling for network issues and temporary server or rate-limit errors (e.g., HTTP 429, 5xx). In these cases, it applies an exponential backoff strategy and enforces a minimum delay between consecutive requests, resuming automatically once the API becomes available again. Non-recoverable errors (i.e., responses with status codes that are not temporary) cause the collection for the current year to stop, with the event recorded in a log file. For each API response, the script extracts a consistent subset of metadata for every model, including: unique identifier, title, author name, author profile URL, publication date, number of views, number of “likes”, number of comments, list of tags, license label, and viewer URL. These records are appended incrementally to a single CSV file, building a consolidated dataset spanning all years considered. Execution can be safely interrupted and resumed. After each page of results, the script stores its state in a JSON file (current year, current page, API cursor, and a running total of collected models), and writes progress messages to a text log. On restart, the script reads this state and continues from the last successfully processed page rather than restarting from the beginning. The complete source code of the data collection script is reported in Appendix A. The API token used for authentication is not hard-coded in the public version of the script; instead, it is read from a local environment variable for security reasons.

2.4. Statistical Analysis and Visualization

Following data acquisition, we conducted a structured descriptive and comparative analysis, summarizing views, likes, and comments with standard descriptive measures (e.g., mean and dispersion) and computing the NMI index for cross-category comparison, to characterize the dataset and identify patterns of content creation and audience engagement. Descriptive statistics were computed to quantify the total number of models and to summarize engagement metrics (views, likes, and comments). In particular, we generated a summary table for the “History & Cultural Heritage” dataset, reporting average (mean) and maximum values for each engagement variable. In addition to these core descriptors, we computed further distributional parameters to better capture the strong skew typically observed in platform metrics, including median, standard deviation, interquartile range (IQR), and selected percentiles (e.g., 25th, 50th, 75th, 90th, 95th, 99th). Where appropriate, log-transformed versions of engagement variables were also considered to reduce the influence of extreme outliers and facilitate comparison across orders of magnitude. To support interpretation and communicate results effectively, we constructed multiple visualization types from the cleaned dataset. Histograms (and complementary kernel density summaries, where needed) were used to represent the distribution of views, likes, and comments; boxplots/violin plots were used to compare engagement distributions across years; line charts quantified temporal trends in annual uploads and engagement over time; and bar charts summarized categorical distributions such as the most frequent tags, the most active authors, and yearly totals. To assess relationships between engagement indicators, we produced scatter plots (e.g., views vs. likes; views vs. comments), optionally using log scales, and computed correlation measures (Pearson/Spearman) to quantify monotonic or linear association. Finally, to capture concentration effects in attention allocation, we derived inequality-oriented summaries (e.g., share of total views attributable to the top x% of models) and, when required, complementary indices for robustness. To identify prolific contributors and high-impact models, the dataset was further interrogated to determine the top 10 authors ranked by the number of published models and the top 10 models ranked by view count. Complementary rankings based on likes and comments were also extracted to distinguish “high-reach” from “high-endorsement” artifacts. All analyses and figures were produced in Python v.3.14.3 (e.g., using standard data-analysis and plotting libraries).

2.5. Approach to Evaluate the Impact the “History & Cultural Heritage” than Other Categories

Complementary visualization techniques, including histograms, bar charts, and scatter plots, were employed to facilitate interpretation and to communicate insights effectively. Collectively, these analyses provide a comprehensive overview of the dynamics of 3D content production and consumption on Sketchfab, situating the study within a broader framework of digital content analysis. To enable a consistent comparison among categories characterized by markedly different magnitudes of user-interaction metrics, we constructed a composite indicator termed NMI (normalized mean index). The label “NMI” is used to emphasise the two defining features of the index (normalization and arithmetic averaging) while the construction follows the established composite-indicator practice based on distance-to-the-best-performer (ratio-to-maximum) normalization and equal-weight additive aggregation [17,18]. In the first stage, each engagement metric x (here n = 4, e.g., count, views, likes, and comments) was normalized by dividing each category value by the maximum observed for the same metric across categories (“best performer”):

$$x_{ij}^{norm}={\displaystyle \frac{x_{ij}}{x_{j,max}}}$$

(1)

where $x_{ij}$ is the observed value of metric j for category i, and $x_{j,max}$ is the maximum value of metric j across all categories. This step rescales each metric to [0, 1] and removes scale disparities, making the metrics directly comparable. Subsequently, for each category i, we computed the unweighted arithmetic mean of the normalized metrics to obtain the NMI score:

$$NM{I}_i={\displaystyle \frac{1}{n}}\sum _{j=1}^n{x}_{ij}^{norm}$$

(2)

The resulting NMI_i ranges from 0 to 1: values approaching 1 indicate that a category performs close to the top across multiple engagement dimensions, whereas values near 0 indicate limited visibility or interaction potential. In addition to descriptive summaries, we performed simple inferential checks: (i) correlation analyses between reach and engagement indicators; and (ii) a log–log linear regression across categories to model how total likes scale with total views and to compute residual-based endorsement intensity.

3. Results

3.1. Quantitative Overview of the Cultural Heritage & History Dataset

After conducting the data extraction from the Sketchfab API, the results revealed a rich and varied dataset. In total, 1,377,107 models were registered, each representing a unique contribution to the platform’s growing digital collection. The dataset was organized across 11 columns (uid, name, author, author_profile, publishedAt, views, likes, comments, tags, license, url) capturing different aspects of each model’s identity and metadata.

Table 1. Statistical overview of all considered models from Sketchfab (“Cultural Heritage & History”, 2018–2025).

Metric	Average	Max
views	332.129	2,974,583
likes	6.682	2963
comments	0.204	97

reports, for each metric, the average value and the maximum value observed across the unified dataset (2018–2025).

To provide a concise yet informative overview of the 3D models analyzed, it is useful to describe the distribution of the main tags associated with the content. Tags act as a direct indicator of both the dominant thematic domains (e.g., archaeology, museum, history) and the most widespread techniques and workflows (e.g., photogrammetry, Metashape, Agisoft, RealityCapture) within the user community. The concentration of occurrences on a limited number of labels also makes it possible to highlight disciplinary and technological biases in the dataset, informing the interpretation of subsequent results (for instance, in terms of thematic coverage, representativeness of specific cultural domains, or dependence on particular software tools). In light of these considerations,

Table 2. Top 10 most frequent tags in the unified Sketchfab cultural heritage dataset (2018–2025).

#	Tag	Count (Models with Tag)	% of All Models
1	photogrammetry	167,093	12.13%
2	archaeology	140,239	10.18%
3	noai	83,627	6.07%
4	museum	68,826	5.00%
5	metashape	68,730	4.99%
6	heritage	61,876	4.49%
7	agisoft	61,695	4.48%
8	realitycapture	60,426	4.39%
9	history	52,939	3.84%
10	medieval	51,380	3.73%

reports the 10 most frequent tags in the unified dataset (2018–2025), together with the number of models associated with each tag and their percentage over the total. This summary offers an immediate overview of the main semantic and technical macro-categories structuring the corpus and provides the reader with essential context for understanding the quantitative and qualitative analyses presented in the following sections.

To characterize the legal conditions under which the models can be reused, we analyzed the distribution of licence types in the Cultural Heritage & History dataset. Figure 2 presents this distribution as a pie chart, grouping all licence categories below 3% into a single “Other” class to improve readability. The chart highlights a strong dominance of records with Unknown licences (≈50.5%), followed by CC Attribution (≈27.2%), CC Attribution-NonCommercial (≈6.1%), Standard (≈5.4%), and CC Attribution-Non-commercial-Share Alike (≈4.0%), with the remaining licences collectively accounting for ≈6.8% of the models. A smaller portion of models was distributed under various other, less common license types, as shown in Figure 2.

To quantify the temporal evolution of the dataset, we analyzed the total number of cultural heritage models published on Sketchfab between 2018 and 2025. Figure 3 shows a marked increase in annual uploads, with the number of models rising from a few tens of thousands in 2018 to more than three hundred thousand in the mid-2020s. A linear regression fitted to the yearly counts (y = 44,999.6 × “year” − 90,794,567.5; R² = 0.96) captures this strong upward trend, while the slight deviation in the most recent year suggests a possible stabilization after a phase of rapid growth.

3.2. Comparison Between the Following Categories

To gain a deeper understanding of the dataset’s internal structure and to identify potential patterns in user engagement, a comparative analysis among the different categories available on Sketchfab was performed. The stacked bar chart (Figure 4) summarizes annual model counts (2018–2025) for each Sketchfab category, with one bar per category and coloured segments representing individual years. This representation highlights both (i) between-category differences in total volume (overall bar height); and (ii) within-category temporal composition (segment sizes).

Across the full observation window, the largest cumulative volumes are concentrated in a subset of categories—most notably weapons-military (10,480 models) and furniture-home (10,142), followed by architecture (9819), art-abstract (9816), fashion-style (9807), and nature-plants (9756). In contrast, news-politics constitutes the smallest overall contribution (4188 models), indicating a comparatively limited presence of this content type in the dataset. Aggregating across categories, the total number of models exhibits a non-monotonic but clearly structured temporal pattern: a decline from 2018 (19,514) to 2020 (16,311), followed by a marked expansion through 2021 (19,859) and 2022–2024, peaking in 2024 (24,342). Many categories mirror this broader increase toward 2024. For example, architecture shows the strongest growth relative to 2020 (approximately +174% from 2020 to 2024) and contributes a substantial 2024 segment (a prominent share of its total). Music and fashion-style also display sustained increases into 2024, consistent with a general broadening of activity during 2021–2024. Some categories deviate from the overall expansion pattern. Furniture-home peaks early (2019: 2366) and then declines sharply, remaining comparatively lower thereafter, suggesting a structural shift in that category’s production (or in how it is captured in the dataset). Weapons-military is highest at the start of the series (2018: 1765), drops in 2019–2020, and then partially rebounds in 2022–2024. Below, we report the NMI for each Sketchfab category over 2018–2025. NMI is computed by first normalizing each category’s yearly count by the mean count across all categories in the same year, then averaging these normalized values across years. In

Table 3. NMI by Sketchfab category (2018–2025).

Category	NMI (Index, 100 = Avg)	Total Models (2018–2025)	Mean per Year
weapons-military	118.9	10,480	1310.0
furniture-home	118.0	10,142	1267.8
art-abstract	110.9	9816	1227.0
nature-plants	110.1	9756	1219.5
architecture	108.9	9819	1227.4
people	108.5	9655	1206.9
fashion-style	108.1	9807	1225.9
music	107.4	9596	1199.5
sports-fitness	104.8	9401	1175.1
food-drink	103.7	9316	1164.5
cultural-heritage-history	100.4	8972	1121.5
cars-vehicles	100.2	8814	1101.8
science-technology	99.6	8950	1118.8
characters-creatures	99.1	8572	1071.5
animals-pets	97.5	8527	1065.9
places-travel	83.8	7277	909.6
electronics-gadgets	73.6	6388	798.5
news-politics	46.6	4188	523.5

, the NMI index is scaled so that 100 = yearly average (values > 100 indicate above-average impact relative to the yearly baseline).

Overall, the NMI distribution indicates a clear stratification of categories in terms of their average, year-normalized contribution. Weapons-military and furniture-home show the strongest relative impact (NMI ≈ 119), meaning that (after controlling for year-to-year fluctuations in overall activity) these categories consistently remain ~18–19% above the yearly category average. A second tier of above-average categories includes art-abstract, nature-plants, architecture, people, fashion-style, and music (NMI ≈ 107–111), suggesting sustained prominence across the period rather than a single-year surge.

Categories with NMI values near 100 (e.g., cultural-heritage-history, cars-vehicles, science-technology, characters-creatures) closely track the annual baseline, implying that their volumes are broadly proportional to the platform-wide trend in each year. In contrast, places-travel and electronics-gadgets are systematically below average (NMI ≈ 84 and 74, respectively), while news-politics is markedly underrepresented (NMI ≈ 47), indicating that its yearly counts are, on average, less than half of the typical category-year baseline. Importantly, because NMI is normalized within each year, it emphasizes relative standing rather than raw volume alone. The inclusion of Total models (2018–2025) helps to contextualize magnitude; however, interpretation of late-period values (notably 2025) should consider potential partial-year coverage if the 2025 counts do not span a full comparable interval. Cultural-heritage-history is very close to the platform average. Its NMI is 1.004 (index 100.4, where 100 = the yearly average), meaning it contributes about the expected amount after accounting for differences in overall activity across years. It ranks in the middle of the categories: lower than the top performers (e.g., weapons-military and furniture-home, ~118–119) but higher than consistently underrepresented categories such as news-politics (~46.6) and electronics-gadgets (~73.6). In absolute terms, it totals 8972 models from 2018–2025 (about 1122 per year), which matches its near-average normalized impact. To explore how visibility and community responses vary across Sketchfab content domains, we distinguish between metrics capturing reach and those capturing engagement at the category level. The results reported in

Table 4. Category-level reach and engagement on Sketchfab.

Category	Total Views (×1000)	Number of Models with >1000 (×1000)	Total Likes (×1000)	Models with >100,000 Likes (Count)
weapons-military	262.5	5	262.5	2
furniture-home	769.7	0	106.5	0
art-abstract	6900.0	2	317.0	4
nature-plants	800.1	0	412.2	1
architecture	7697.0	110	509.7	15
people	2200.0	2	267.9	8
fashion-style	11,600.0	1	241.4	1
music	77.9	0	28.7	0
sports-fitness	294.1	0	68.3	0
food-drink	140.9	0	41.8	0
cultural-heritage-history	3000.0	4	787.7	5
cars-vehicles	2200.0	4	751.7	8
science-technology	1700.0	2	787	11
characters-creatures	8900.0	6	558.4	17
animals-pets	16,000.0	5	245.3	2
places-travel	509.7	0	509.7	13
electronics-gadgets	2300.0	1	83.4	0
news-politics	2200.0	3	98.7	0

can be interpreted as evidence that category-level performance on Sketchfab is shaped by two partially distinct dimensions, reach and engagement, which do not necessarily move together. The table operationalizes reach through total views (in millions) and the number of models surpassing 1M views, while engagement is proxied by total likes (in thousands) and the number of models exceeding 100k likes. Read in combination, these indicators suggest that some categories function as high-traffic domains that attract broad, often rapid consumption, whereas others generate comparatively fewer impressions but elicit stronger community endorsement, consistent with more “value-dense” attention.

Read in combination, these indicators suggest that some categories function as high-traffic domains that attract broad, often rapid consumption, whereas others generate comparatively fewer impressions but elicit stronger community endorsement, consistent with more “value-dense” attention.

An inferential check at the category level showed that, after log-transforming totals to account for heavy-tailed distributions, total views and total likes are positively correlated (Pearson r = 0.57, p = 0.013, n = 18 categories). A log–log regression of likes on views indicates sublinear scaling (β = 0.38, p = 0.013, R² = 0.33). Cultural Heritage & History shows a positive residual (rank 4/18), corresponding to ~2.7× more likes than predicted from its view volume, consistent with comparatively “value-dense” attention. In terms of reach, the distribution is markedly skewed. A small group of categories captures the largest total audiences, led by animals-pets (16.0M views) and fashion-style (11.6M), followed by characters-creatures (8.9M) and architecture (7.697M). These categories plausibly benefit from broad, immediate visual appeal and high discoverability, which can translate into repeated browsing and sharing. At the other end of the spectrum, music (0.0779M), food-drink (0.1409M), and sports-fitness (0.2941M) exhibit limited reach and show no evidence of breakout models at either threshold, suggesting that, within this dataset, these domains are either smaller in catalogue size, less frequently searched, or less aligned with the platform’s dominant consumption patterns. A particularly salient result is the exceptional behaviour of architecture when breakout visibility is considered. Architecture reports 110 models exceeding 1M views, far surpassing all other categories. This implies a category environment where a large number of artifacts consistently reach high visibility, which may reflect structural factors such as utility in professional workflows, educational reuse, embedding in external sites, or sustained long-tail discovery. The same category also shows strong engagement in the upper tail (15 models > 100k likes), indicating that high visibility is accompanied by a substantial subset of highly endorsed models. However, the defining feature remains the sheer volume of ultra-high-view content, suggesting that architecture is not merely producing a few viral outliers but rather a broad base of frequently consumed assets. Against this landscape, cultural-heritage-history emerges as one of the most consequential categories when impact is defined not only as exposure but as the capacity to elicit endorsement. Cultural Heritage & History achieves 3.0M views, which positions it below the high-traffic categories that dominate reach (e.g., animals-pets, fashion-style, characters-creatures, architecture). Nonetheless, it records 787.7k likes, placing it at the highest tier of engagement in the table and effectively matching science-technology (787k likes) while exceeding cars-vehicles (751.7k) and characters-creatures (558.4k). This divergence between moderate reach and exceptional likes suggests that cultural heritage content may attract a form of attention that is qualitatively different—more intentional, appreciative, or value-driven—consistent with the educational and cultural significance often associated with heritage assets. The threshold-based indicators further support this interpretation while also delineating the nature of Cultural Heritage’s impact. The category reports four models > 1M views, demonstrating that it is capable of producing widely viewed “flagship” items, even if it does not generate them at the scale of architecture. At the same time, it shows five models > 100k likes, indicating that strong endorsement is not limited to a single standout artifact. Importantly, Cultural Heritage does not lead the table in the number of highly liked models (characters-creatures (17), architecture (15), and places-travel (13) show broader upper-tail density by this measure) suggesting that Cultural Heritage’s engagement advantage is driven more by high aggregate likes and a select subset of strongly endorsed models, rather than by a very large number of separate models repeatedly crossing the 100k like threshold. Comparisons with high-view categories illustrate the practical meaning of this pattern. Animals-pets, despite being the most viewed category (16.0M views), accumulates 245.3k likes and only two models > 100k likes, whereas cultural-heritage-history attains substantially fewer views (3.0M) but far higher likes (787.7k) and more highly liked models (5). This contrast indicates that reach alone can reflect broad but potentially low-commitment consumption, while likes may capture deeper approval. Similarly, fashion-style combines very high views (11.6M) with comparatively modest likes (241.4k) and only one model > 100k likes, reinforcing the idea that some categories are optimized for exposure but yield weaker engagement intensity. In this sense, Cultural Heritage & History appears to “convert” impressions into endorsement more effectively than several categories that outperform it in raw audience size. The closeness between Cultural Heritage & History and Science & Technology is noteworthy. Both accumulate roughly 787k likes, but Science & Technology records fewer views (about 1.7M) while featuring more highly liked models (11 with >100k likes, versus fewer in Cultural Heritage & History). This suggests different engagement patterns: Science & Technology appears to spread strong endorsement across a broader set of models, whereas Cultural Heritage & History concentrates very high appreciation in a smaller upper tail, while still producing a non-trivial number of breakout items. Cars & Vehicles further nuances the comparison, pairing moderate views (about 2.2M) with very high likes (751.7k) and eight models above 100k likes, consistent with the idea that categories centred on recognizable real-world objects and enthusiast communities can generate high-intensity engagement. Finally, the interpretation of these patterns should acknowledge limitations in the provided data. Some cells are missing (e.g., furniture-home lacks threshold counts), and at least one entry (e.g., “317k” for art-abstract under a “Like (×1000)” header) appears formatted inconsistently with the rest of the column, potentially reflecting a transcription or unit standardization issue. These constraints do not negate the directional conclusions drawn from the largest contrasts (particularly the reach dominance of animals-pets and fashion-style, the breakout volume of architecture, and the engagement prominence of cultural-heritage-history); however, they imply that any formal inference (e.g., normalized rates, cross-category effect sizes, or statistical tests) would require cleaning and harmonizing units before proceeding.

4. Discussion

4.1. Explaining Engagement Patterns in History & Cultural Heritage

The analysis of the History & Cultural Heritage category on Sketchfab reveals a rapidly expanding and increasingly structured ecosystem of 3D content. Between 2018 and 2025 the number of published models grows from roughly 26,000 to more than 300,000 items per year, with a strong linear trend and only a slight decline in the most recent year. This pattern suggests that 3D digitization has become a consolidated practice in the cultural heritage domain rather than a short-lived experimentation phase. The continuous growth also indicates that the platform is progressively being adopted as a standard repository for sharing and disseminating 3D heritage data, both by institutions and by individual practitioners. The tag analysis highlights a clear thematic and technological bias in the corpus. Tags such as photogrammetry, archaeology, museum, heritage and history dominate the top positions, together with software-specific labels such as Metashape, Agisoft, and RealityCapture. This concentration demonstrates that the dataset is strongly oriented towards archaeological and museum contexts and that image-based modelling is the prevailing acquisition technique. While this is consistent with current professional practices, it also implies limited representation of other kinds of heritage (e.g., intangible heritage, contemporary architecture, industrial or landscape heritage) and of alternative digitization workflows (e.g., laser scanning, procedural or hand-modelled reconstructions). Consequently, results derived from this corpus should be interpreted as primarily characterizing photogrammetric, archaeology-driven digitization. User engagement metrics further nuance this picture. On average, models receive a few hundred views but only a small number of likes and very few comments. This suggests that, although the platform attracts considerable visual traffic, active interaction and discussion around individual models remain relatively limited. The long tail of models with zero comments and very few likes may indicate a primarily “consumptive” use of Sketchfab, in which 3D content is viewed or embedded but not necessarily discussed or collaboratively curated. For cultural heritage professionals, this raises questions about how effectively 3D models function as tools for dialogue, learning and community building, rather than as static digital surrogates. Licensing patterns have important implications for data reuse. Approximately half of the models present an Unknown licence, while most of the remaining content is distributed under Creative Commons licences, with CC Attribution as the dominant explicit choice. On the one hand, the strong presence of CC licences confirms that a substantial fraction of heritage models is intended to be reusable, at least in non-restrictive academic or educational contexts. On the other hand, the prevalence of Unknown licences introduces significant legal uncertainty: models without clear reuse terms cannot be safely integrated into open pipelines, aggregated into derived datasets, or redistributed in new platforms. The presence of non-commercial and share-alike clauses in a non-negligible subset of the corpus further complicates reuse for mixed or commercial projects.

From a heritage-data governance perspective, this underlines the need for clearer licensing workflows and for awareness-raising among contributors. Taken together, these findings portray a vibrant but uneven digital heritage landscape. The category under study is characterized by rapid quantitative growth and by strong alignment with current photogrammetric practice; however, it also exhibits topical and technological imbalances, limited user interaction and significant ambiguity regarding licensing. For researchers, this means that Sketchfab can serve as a rich empirical laboratory for large-scale studies on 3D heritage production and dissemination, provided that analyses explicitly acknowledge these biases. For institutions and policymakers, the results argue for targeted actions to diversify the types of documented heritage, promote clearer and more open licensing choices, and foster forms of engagement that go beyond mere viewing. To complement the qualitative comparison between repositories, we also inspected (where platform taxonomies make this possible) the approximate “nearest equivalent” of Sketchfab’s Cultural Heritage & History category on other major 3D platforms. Market-oriented repositories typically expose heritage content through pragmatic, use-case labels (e.g., “archaeology” props or “historic” environments), rather than through an explicit cultural-heritage framing. For instance, CGTrader lists 1734 “Archaeology” models, and TurboSquid lists 2461 “archaeology” models, both explicitly presented as downloadable assets for games/VR/AR pipelines. By contrast, the Unity Asset Store surfaces heritage primarily as an environment genre: its “3D Historic” page reports 1–24 of 827 results, reinforcing the idea that “heritage” is operationalized as a production setting (historic streets, temples, ruins) rather than as a documentation-driven corpus of specific artifacts. When placed next to our Sketchfab evidence (where Cultural Heritage & History achieves high interaction intensity relative to its views, with engagement often concentrating on institutionally curated uploads), these inventories suggest a structural divergence: marketplaces concentrate heritage into a comparatively small number of commercial packs and generic props, whereas open repositories can accumulate a substantially broader long tail of digitized items and museum-driven scans.

Table 5. Cross-platform proxy counts for heritage-related 3D content.

Platform	Closest Available Taxonomy (Proxy)	Approx. Count
Sketchfab	Cultural Heritage & History (category corpus in our dataset)	1,377,107
CGTrader	Archaeology (category page)	1734
TurboSquid	Archaeology (category page)	2461
Unity Asset Store	3D Historic (Environments)	827

summarizes a cross-platform estimate of the size of “heritage-related” 3D inventories by using the closest publicly available category on each platform as a proxy for Sketchfab’s Cultural Heritage & History. Because platforms adopt different taxonomies and serve different audiences (e.g., public dissemination vs. commercial marketplaces), categories are not perfectly comparable. For this reason, Table 5 should be read as an indicative benchmark of how explicitly “heritage-coded” content is surfaced and quantifiable through platform navigation, rather than as a strict equivalence of cultural heritage corpora. Specifically, Sketchfab offers an explicit Cultural Heritage & History category, whereas commercial marketplaces tend to expose heritage mainly through production-oriented labels such as “Archaeology” (CGTrader and TurboSquid) or “Historic” environment packs (Unity Asset Store). The counts reported correspond to the publicly visible number of items returned under each proxy label at the time of inspection, and they therefore capture both (i) the scale of inventory; and (ii) how heritage content is operationalized within each platform’s browsing structure.

The disparity is visually apparent in Figure 5, where a log scale is required to place repository-style accumulation (Sketchfab) alongside marketplace-style inventories (CGTrader, TurboSquid, Unity).

Beyond counts, the comparison also highlights a metrics asymmetry that affects how “impact” can be measured. Sketchfab foregrounds social attention (views, likes, and highly viral outliers), while most marketplaces foreground transactional discovery, where the most relevant signals (sales, conversions, pipeline adoption) are typically not publicly comparable. In this context, our interpretation of “impact” for cultural heritage differs by platform: on Sketchfab, impact is visible as concentrated engagement around culturally “canonical” objects, and our qualitative inspection aligns with a recurrent pattern in which the most viewed and most interacted-with heritage items are tied to famous artworks by renowned artists and/or digitizations associated with major museums (i.e., strong institutional brands functioning as attention hubs).

On marketplace platforms, however, the closest analogues to “cultural heritage” are frequently optimized for re-use (ruins kits, historic props, period environments) rather than for scholarly provenance; therefore, impact tends to be expressed through production utility rather than through public-facing interaction metrics. In addition, two additional reference points help to contextualize why “heritage-like” assets may appear differently across ecosystems. ArtStation Marketplace is large at the level of overall game assets (e.g., tens of thousands of “Game Assets” results), but does not consistently isolate “cultural heritage” as a stable category, making direct quantification difficult. Superhive similarly documents overall marketplace scale (e.g., 49,117 model products and 59,603 total products); however, its public statistics are not broken down into a heritage-specific taxonomy. Meanwhile, Quixel’s ecosystem—now closely tied to Epic’s broader marketplace strategy—frames impact through a continuously expanding scan library; Quixel notes “over 18,000 assets and counting” available via Fab for Megascans/Megaplants access. Fab’s transition documentation notes that 544 legacy Megascans assets were deprecated for quality reasons, illustrating active curation pressures even in very large libraries.

4.2. Sketchfab in the Context of European Cultural Heritage Digitization

These results must be considered within the broader framework of the European Commission Recommendation on a Common European Data Space for Cultural Heritage (C (2021) 7953 final, 10 November 2021) [19]. The Recommendation emphasises that advanced digital technologies—particularly 3D digitization, virtual and augmented reality, and cloud infrastructures—represent key enablers for accessibility, reuse, and the preservation of Europe’s cultural assets. It calls upon Member States to digitize all monuments and sites at risk by 2030 and at least 50% of major monuments and sites by 2025, while promoting open access, interoperability, and the reuse of digitized content through Europeana and the future common data space. Within this policy context, Sketchfab functions as a crucial complementary platform that operationalizes many of these recommendations. By enabling institutions and creators to upload, annotate, and share high-quality 3D models of cultural artefacts, Sketchfab facilitates the public dissemination and reuse of digitized cultural heritage. Its web-based interface and embedding functionalities lower technical barriers for access, while its compatibility with open standards (e.g., glTF, WebGL) aligns with the European emphasis on FAIR (Findable, Accessible, Interoperable, and Reusable) principles. However, the empirical findings from this study demonstrate that the potential of Sketchfab remains under-realized in terms of user engagement for heritage content. While the platform’s infrastructure supports the technological requirements identified in the EU Recommendation, such as 3D digitization and open online accessibility, the communicative and participatory dimensions of cultural heritage remain weakly developed. Engagement levels are often hindered by a lack of narrative contextualization, pedagogical framing, and emotional mediation that would enable non-expert users to connect meaningfully with digital artefacts.

Moreover, Sketchfab’s community dynamics and visibility algorithms tend to favour entertainment-related content (e.g., gaming, design, animation), thereby limiting exposure for educational and cultural heritage models. This structural asymmetry indicates that, while Sketchfab aligns with EU goals for digitization and access, it only partially fulfills the broader objective of fostering active public participation and co-creation, core aspects of the digital transformation envisioned by the European Commission (Recital 8 and Chapter II, Articles 4–9 of the Recommendation) [19]. Indeed, Sketchfab’s model offers strong alignment with EU policy priorities in several respects:

• it contributes to 3D digitization and reuse of cultural heritage assets, directly supporting the 2030 targets for 3D documentation;
• it promotes cross-sectoral collaboration, engaging museums, designers, researchers, and educators, consistent with the Recommendation’s call for partnerships between the cultural heritage sector and creative industries;
• its cloud-based and interoperable nature supports the development of the Common European Data Space for Cultural Heritage, facilitating data sharing across institutional and national boundaries;

Nevertheless, several limitations must be addressed for full policy integration:

• lack of metadata standardization and variable licensing frameworks complicate interoperability with Europeana and the forthcoming data space;
• insufficient digital preservation strategies risk compromising the long-term accessibility of Sketchfab-hosted assets, a key requirement under Chapter II, Article 6 of the Recommendation [19];
• limited professional digital skills among heritage practitioners hinder optimal exploitation of 3D and immersive technologies, reflecting the skills gap identified in Recital 15.

Therefore, to maximize its contribution to the EU’s vision of a digitally transformed and resilient cultural heritage sector, Sketchfab could evolve from a generalist 3D sharing platform to an integrated hub of the European heritage data ecosystem. This would require a more structured collaboration with Europeana and the adoption of common metadata and rights frameworks (Europeana Data Model, RightsStatements.org), as well as the development of heritage-specific curation tools capable of enhancing storytelling, accessibility, and participation.

Over the past decade, the European digital heritage ecosystem has gradually shifted from a logic centered on 2D aggregation to a model in which 3D assets are considered primary objects of research and dissemination, with Europeana increasingly playing the role of public access layer for discovery, contextualization, and reuse. Along this path, several EU-funded initiatives demonstrate how 3D digitization, interoperability, and long-term access have been operationally implemented. A key step was the CARARE best practice network (2010–2013), aimed at integrating archaeological and architectural content into Europeana on a large scale [20]. Beyond aggregation, CARARE defined workflows and mappings that enabled the consistent ingestion of place-based (often georeferenced) records and related digital assets, including 3D/VR materials, strengthening interoperability between domain repositories and the Europeana infrastructure. Building on this, 3D-ICONS (2012–2015) promoted a more explicitly productive agenda, aiming at the digitization of architectural and archaeological “icons” and the publication of 3D models and related content in Europeana, contributing to the formation of a critical mass of 3D resources with high engagement potential [21]. Equally importantly, the project has documented end-to-end practices (acquisition, processing, metadata, publication) through guidelines and case studies, fostering methodological transferability beyond the consortium. Within this framework, 3D models are not produced solely for visualization, but are treated as structured assets designed for interoperability, preservation, and downstream reuse, consistent with European efforts to scale access and utility for institutions and audiences. Overall, these Europeana-related initiatives highlight the value of an approach that integrates high-quality 3D digitization, standards-based interoperability, and sustainable dissemination, thus maximizing the scientific value, accessibility, and long-term reusability of digital cultural heritage assets. More recently, other initiatives such as LIP3D (Living forever the Past through a 3Digital world) utilized also the CARERE platform reflecting the shift toward integrated pipelines that connect 3D digitization with extended reality (XR) and the emerging Data Space for Cultural Heritage paradigm [22]. Figure 6 offers a concrete example: the Europeana “Overview (MASTERPLAN) of the Archaeological Park of Iuvanum” card presents a navigable masterplan via an interactive viewer and is enriched with hotspots/annotations.

4.3. Strategies to Improve Cultural Heritage Models

The evidence discussed above indicates that Cultural Heritage & History combines comparatively moderate reach with exceptionally strong engagement, suggesting that the category’s core limitation is not audience appreciation but rather discoverability and scaling of exposure. In practical terms, strategies should therefore prioritize increasing the probability that high-quality heritage assets are encountered by broader audiences, while preserving the interpretive richness that likely drives the category’s unusually high like counts. This orientation is consistent with the platform’s longstanding institutional ecosystem: Sketchfab has explicitly documented the growth of cultural-institution participation and noted that cultural-heritage uploads can reach very high view counts, including a cultural-heritage scene reported as exceeding 1.4M views and ranking among the platform’s most viewed scenes [23]. A second implication concerns what tends to travel widely once it is surfaced. Platform-curated museum-focused collections and institutional profiles suggest that highly visible heritage content is frequently associated with recognized museums and large cultural organizations, which can function as credibility and traffic multipliers. In parallel, the circulation of “canonical” artworks appears to operate as an attention anchor: dedicated collections explicitly framed around rediscovering world-famous paintings in 3D, together with multiple user-uploaded Mona Lisa reconstructions, illustrate how globally recognizable works and renowned authorship can lower the cognitive barrier to engagement and sharing. This pattern does not imply that lesser-known artifacts lack value; rather, it suggests that recognizability, whether intrinsic (iconic works) or constructed (through framing and context), is a major driver of reach. Accordingly, improving performance in Cultural Heritage & History can be approached as a combination of editorial, technical, and distribution interventions. On the editorial side, models benefit from being published as part of coherent narratives and curated sequences rather than as isolated objects, because narrative framing effectively “manufactures recognizability” for non-iconic items by explaining why they matter (provenance, function, dating, cultural context, restoration history). On the technical side, raising baseline presentation quality improves first-impression retention—particularly through consistent lighting, accurate scale, clean materials, and carefully tuned viewer settings, which is important, given that users often decide within seconds whether to continue interacting. On the distribution side, discoverability is strengthened by disciplined metadata practices (standardized titles, rich descriptions, multilingual keywords, where relevant, and tags that connect heritage assets to adjacent high-traffic discovery paths such as architecture, travel, and art), and by intentional integration with external channels that generate sustained attention (museum websites, educational portals, exhibition pages, and press releases). Finally, where possible, institutional collaborations should be treated as a strategic lever rather than merely an attribution detail: museum partnerships and collection-level publishing can transfer trust, visibility, and cross-promotion capacity to individual models, reinforcing the category’s demonstrated ability to convert exposure into endorsement once audiences are reached.

5. Conclusions

This study mapped the Cultural Heritage & History ecosystem on Sketchfab through a reproducible, API-driven workflow and a large-scale analysis of public model metadata. Across 1,377,107 models, the results show a domain in rapid expansion over 2018–2025, but also a strongly unequal attention economy: most heritage models attract limited interaction, while a small minority concentrates a disproportionate share of views and likes.

The tag landscape indicates a corpus shaped by prevailing digitization practices, with photogrammetry and archaeology/museum terminology dominating the descriptive vocabulary. At the same time, reuse conditions remain substantially constrained by rights uncertainty: approximately half of the records are associated with an Unknown licence, limiting safe downstream aggregation and reuse. These patterns underscore the importance of clearer rights workflows and richer metadata/paradata if 3D heritage is to support robust research, education, and cross-platform interoperability. When compared with other Sketchfab categories, Cultural Heritage & History sits close to the platform’s year-normalized baseline in volume (NMI ≈ 100). However, it exhibits comparatively high endorsement relative to reach, consistent with “high-value” engagement once heritage content is discovered. This suggests that the principal constraint for heritage models is less about audience appreciation than discoverability and contextualization—areas where improved storytelling, curatorial framing, and integration with institutional channels (including Europeana-aligned infrastructures) can plausibly increase reach without sacrificing interpretive depth. Beyond providing a descriptive snapshot, this baseline is important because it makes a largely opaque platform ecosystem measurable and comparable. It offers institutions, curators, and researchers an evidence base for decisions about digitization priorities, publication strategies, and audience development, while also revealing where current practices fall short—most notably in licensing clarity and contextual metadata. As such, the study helps to shift 3D heritage publishing from ad hoc dissemination toward more accountable, reusable, and interoperable digital scholarship. Methodologically, the work contributes a transparent extraction and analysis pipeline designed to cope with cursor-based pagination, intermittent failures, and rate limiting, and documents how large language models can assist in scripting and verification tasks when their outputs are critically reviewed by domain experts. Limitations include reliance on platform-provided metadata (including possible category misassignment and incomplete provenance), the snapshot nature of engagement metrics, and the imperfect comparability of cross-platform “heritage” proxies. Building on the baseline established here, future research can: (i) monitor the ecosystem longitudinally and test the effects of curatorial or metadata interventions on discovery and engagement; (ii) investigate how recommendation and editorial curation shape visibility and bias; (iii) develop tag normalization and entity-linking methods to connect models to controlled vocabularies and knowledge graphs; and (iv) strengthen provenance and rights workflows by reconciling Unknown licences and aligning platform records with Europeana/EDM-oriented metadata and institutional documentation through targeted case studies. Together, these directions position platform-scale analytics as a practical foundation for more sustainable, transparent, and reusable 3D heritage infrastructures.