Daejeon, July 27–31
1. Introduction
This research is part of the Narod.ru and Digital Heritage project (Aslanov et al. 2025), which preserves and studies websites of Narod.ru – a free hosting service active from 2003 to 2013. Narod.ru was a free web hosting service launched in 2000 and widely used in Russian‑speaking and post‑Soviet communities. It allowed users to create personal websites without technical skills, becoming a significant platform for early digital self‑expression, amateur art, fan sites, and grassroots publishing. Although based in Russia, it was used not only by Russian speakers but also in several other countries. The period of its greatest activity, as identified by the authors, was from 2003 to 2013. Its popularity declined after the service was sold in 2013 to a new owner and with the rise of social networks. The authors classify Narod.ru hosting as an early web phenomenon (Web 1.0), since its popularity and sociocultural role were formed in the period preceding the dominance of Web 2.0 ideology and mass social networks. UNESCO. Charter on the Preservation of Digital Heritage. 15 October 2003. <https://www.unesco.org/en/legal-affairs/charter-preservation-digital-heritage> [10.12.2025]. One terabyte of kilobyte age, Tumblr. <https://oneterabyteofkilobyteage.tumblr.com> [10.12.2025].
A major initiative to rescue disappearing sites is the Wayback Machine (Internet Archive). However, for a sample of 10,000 Narod.ru sites it contains snapshots of only 71.3% of homepages and 22.3% of subpages. Therefore, one of the core goals of the Narod.ru and Digital Heritage project is to preserve websites comprehensively rather than rely on partial archival captures (Aslanov et al. 2025).
Beyond preservation, the project aims to make this large and heterogeneous corpus analytically accessible by applying quantitative methods, interpreting aggregate patterns, selecting representative case studies, and developing approaches to structuring preserved web data.
In this paper, we focus on one of these analytical challenges and develop a methodology for extracting authors’ declared motivations for publishing art-related materials. We test this methodology on preserved Narod.ru websites assigned to the Visual Art & Design category. By motivation we mean declared reasons, goals, and expectations that led creators to sharing works, posting their own artworks, and organizing website content. Motivation is expressed through explicit author statements and through site structure (e.g., features for uploading artworks). It is reconstructed from digital traces and does not capture hidden motivations.
As the main tool for extracting motivations, this work uses large language models (LLMs). Modern LLMs are increasingly applied to modeling and analysing human behavior, including aspects of Theory of Mind relevant to motivation studies (Zhuang et al. 2024, Peters / Matz 2024, Xie et al. 2025). In our study, LLMs enable identification and interpretation of site authors’ motivations at a scale where manual annotation would be impractical.
The main contributions are:
2. Thematic classification of websites
To identify authors’ declared motivations for creating visual art-related websites, we first extracted thematically relevant sites from the full dataset. The project uses Google’s topic taxonomy (Google Cloud 2025), which includes 27 top-level categories and the Visual Art & Design subcategory within Arts & Entertainment.
During the entire project, researchers retained a random sample of 10,000 websites hosted on Narod.ru. The study used texts from preserved pages, while site topics were determined from homepage texts. Before extracting Visual Art & Design sites, we first identified sites in the broader Arts & Entertainment class. Because processing the entire corpus with proprietary LLMs is costly, we used locally deployed open models framed as binary classifiers.
To verify classifier performance, we manually annotated randomly selected 400 websites, which provides sufficient reliability for estimating a proportion with a 95% confidence interval and ±5 percentage-point error (Lohr, 2019). Two annotators labeled sites for affiliation with the Arts & Entertainment category. Initial agreement measured by Krippendorff’s alpha (Krippendorff 2011) was 0.68; disagreements were resolved through joint discussion, resulting in full agreement. Ultimately, 50 of the 400 sites were assigned to the target class.
For classification, we tested the following open models available on our server with four RTX5090 GPUs: Qwen3-Next-80B-A3B-Instruct-FP8 (Qwen Team 2025b), Qwen3-30B-A3B-Instruct-2507 (Yang et al. 2025), Qwen3-4B-Instruct (Yang et al. 2025), QwQ-32B (Qwen Team 2025a), and Gemma-3-27b-it (Kamath et al. 2025). The models were prompted to decide whether each homepage belonged to Arts & Entertainment. The best result was achieved by QwQ-32B (macro-F1 = 0.897). Prompts and detailed results are available in the repository Repository with supplementary materials <https://github.com/erwyn-montgomery/adho2026> [05.05.2026].
QwQ-32B was then applied to the full corpus and identified 1,219 Arts & Entertainment sites. From these, Visual Art & Design sites were extracted: the model classified 235 sites into this category, and manual validation confirmed 218 of them (92.8%).
3. Extracting motivation with LLMs
To extract website creators’ motivations, we concatenated texts from all pages of each website into a single document. Motivation extraction followed the text clustering as text classification approach (Huang, He 2024) and consisted of three stages: (1) collecting all potential motivation types; (2) aggregating these types into broader categories using an LLM; and (3) classifying websites into the resulting categories.
In total, we identified 74 motivation types, along with a separate category Not Enough Data to Extract Motivation. These types were aggregated using proprietary models gpt-5.1-high OpenAI. Introducing GPT-5.1 for developers. 13 November 2025. <https://openai.com/index/gpt-5-1-for-developers> [10.12.2025]. Google Cloud. “Gemini 3 Pro.” Google Cloud Documentation: Generative AI on Vertex AI. Last updated 14 December 2025. <https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro> [10.12.2025]. Anthropic. Introducing Claude Opus 4.5. 24 November 2025. <https://www.anthropic.com/news/claude-opus-4-5> [10.12.2025]. LLM Leaderboard – Best & Chat AI Models Compered <https://lmarena.ai/leaderboard/text> [10.12.2025]. Openrouter <https://openrouter.ai> [10.12.2025]. Ibid
At the final stage, we applied 17 binary classifiers, one per motivation category. The LLM classification results were manually validated by the authors. After validation, the model correctly identified motivations for 92.7% of websites (202/218). Two categories (Copyright Protection and Information Resources about Animals) were identified as LLM errors: websites mentioned these topics but were not primarily devoted to them. The multi-label classification results for 202 websites are shown in Figure 1.
Figure 1. Distribution of websites across motivation categories
The most frequent motivation category is Personal Creativity and Self-Expression. Such sites primarily showcase personal artwork and are often combined with motivations to preserve works, sell them, or accept commissions. In cases involving commercial services, authors typically present portfolios to attract clients. Additional details are available in the repository Ibid
4. Case analysis
We examine two websites assigned by the model to Personal Creativity and Self-Expression and Archiving and Documenting Creative Work to illustrate how digital traces support motivation inference.
4.1. Case: Curation to legitimize creativity
The website Paintings and Miniatures from Landscape Stones The websites presented here and below are cited in an anonymized form for ethical reasons. Homepage text: “Hi! Did you come for a sensation? That’s hard these days. This site is no exception. And yet, come in. Perhaps you, too, will find something you have never seen before. And if your eyes open wide at what you see, then I have managed to surprise you. Good luck!”.
4.2. Case: Grassroots curation as a basis for community
The second case illustrates a hybrid motivation in which personal creativity underpins an educational project and community formation. The website Fiction and Fantasy was assigned the categories Archiving and Documenting Creative Work, Implementation of Cultural Projects, Personal Creativity and Self-Expression, Community Building and Fan Activities, and Popularization of Artists and Art Movements. These categories are reflected in the site’s structure and content Website texts: “I just hope that what I like will be interesting to you as well,” “Biographies and bibliographies of the best (in my view) fiction writers,” “as I hope, many art galleries <...> created by you will be posted here,” “perhaps new Harrisons and Vallejos will begin their creative path from here?”, “…I dare assure you that in those rare minutes when laziness does not win, I work hard on what I promised.” “…say a few words about the site author, if he is worthy of your precious attention, and finally just chat on completely free topics, and acquire, if not like-minded people, then pen pals.”
The author’s main goal is to create a shared space for contemporary authors and introduce readers to works within a selected genre. At the same time, the author assumes the role of curator, selecting works based on personal taste and taking responsibility toward users. Hosting a library and a gallery supports assignment to additional categories. The motivations form a hierarchy: personal passion (self-expression) enables systematization and popularization, which in turn make community formation possible. The site becomes a digital third place organized around shared cultural interest rather than commerce or formal membership.
5. Conclusion
Labeling websites by topics and motivations enriches digital heritage objects and supports interdisciplinary research by enabling preliminary assessment, targeted sampling, and more efficient discovery of case studies, as our two cases demonstrate. In this study, LLMs identified Visual Art & Design websites and were then used to analyze authors’ declared motivations. The models demonstrated high effectiveness in both thematic classification (accuracy = 92.8%) and motivation extraction (accuracy = 92.7%), as confirmed by manual validation.
Modern LLMs can substantially reduce annotation effort while addressing questions of human motivation at scale. Future work may examine these results through the lens of participatory culture and fan curatorship (Jenkins 2006) focusing on amateurs’ practices of selection, classification, and public representation of digital objects within the early web’s limited technological environment. Our methodology also engages with broader discussions in web archive research, where scholars have identified similar challenges in corpus definition and curatorial infrastructure (Maemura et al. 2021). The proposed methodology can scale to other website archives and contribute to broader research on digital heritage.