In-Browser Agentic Web: a Decentralized Approach to Information Access

IN-BROWSER AGENTIC WEB: A DECENTRALIZED APPROACH TO INFORMATION ACCESS S. Zerhoudi, M. Granitzer, University of Passau, Passau, Germany Abstract The centralization of web search raises critical concerns regarding privacy protection and user autonomy in information access. While advancements in web technologies offer new possibilities for personal information management, current search systems typically process user data on external servers with limited personalization options. This paper introduces a conceptual methodology for browser-based web indexing that processes and stores data locally, addressing these privacy and control limitations. Our approach implements targeted crawling mechanisms aligned with individual user interests and maintains all operations within the browser environment. The technical framework converts web content into dense vector representations through semantic embedding techniques, enabling efficient storage and retrieval within browser constraints. The architecture features: (1) an in-browser language model for semantic search and context-aware content generation, (2) adaptive crawling algorithms that adjust parameters based on storage limitations and user behavior, and (3) incremental updating mechanisms to maintain index freshness. Evaluation approaches using both simulation-based assessment and human participant validation are proposed. This work encourages research on privacy-preserving web search technologies and establishes a foundation for developing user-controlled information retrieval systems. INTRODUCTION The digital ecosystem’s rapid growth in web content creates both opportunities and challenges for information retrieval. Current web search services, controlled by a few major corporations like Google and Microsoft, typically employ user tracking, centralized indexing, and undisclosed algorithmic methods, raising concerns about privacy, data sovereignty, algorithmic transparency, and offline access [Granitzer et al.(2024),Hendriksen et al.(2024a)]. These centralized search providers rely on collecting and analyzing user data to improve search relevance and advertising revenue, which raises ethical questions regarding privacy and control. Their algorithmic processes often lack transparency, potentially enabling manipulation or biased results influenced by commercial or political factors [Granitzer et al . (2024)]. This opacity undermines user trust and may compromise information reliability. Additionally, the market dominance of a few search providers has also led to practices like Search Engine Optimization (SEO), where content creators prioritize algorithmic visibility over informational quality and user value. In response to these limitations, research interest has shifted to decentralized, transparent, and privacy-conscious alternatives. The Open Web Index (OWI) initiative promotes openly accessible indexing infrastructures and standards, emphasizing transparency, collaboration, and open data principles [Hendriksen et al . (2024b)]. OWI addresses centralized indexing challenges by creating public data structures that democratize search engine development. This project employs extensive indexing operations supported by highperformance computing (HPC) resources across Europe, aiming to diversify the digital information ecosystem. Concurrent with these large-scale efforts, advances in browser-based AI inference technologies have created new possibilities for privacy-focused and personalized web indexing. Recent developments, such as WebLLM [Ruan et al . (2024)], demonstrate the feasibility of running sophisticated AI models directly within browsers. These technologies leverage WebGPU [Kenwright(2022)] and WebAssembly [Haas et al . (2017)] to enable efficient local processing without external cloud services. By processing data locally, these browser-based approaches inherently enhance privacy and user autonomy. Adaptive web crawling techniques driven by semantic modeling of user interests have emerged as essential components for personalized retrieval [Durga et al . (2024)]. Unlike traditional fixed crawling algorithms, user modeling approaches that adapt to browsing patterns and real-time interactions improve retrieval accuracy and relevance. This adaptive methodology ensures content remains specific and current, enhancing user experience. Our research proposes an in-browser web indexing approach that integrates targeted, adaptive crawling and content acquisition based on user-defined interests, local indexing using compressed vector embeddings, and semantic search powered by browser-based language models. This methodology addresses limitations of centralized systems by prioritizing privacy, personalization, offline functionality, and user control. The approach centers on creating a localized, browsercontained semantic index using compressed dense embeddings, providing contextual understanding beyond keywordbased techniques. This allows the system to deliver personalized search results within the user’s local environment. Our work extends principles from the OWI initiative but adapts them to browser environments. Rather than employing collaborative indexing at scale, our approach focuses on localized data organization, efficient embedding methods, and streamlined inference capabilities suitable for resourcelimited personal computing contexts. Our contributions include: (1) proposing a conceptual design for a decentralized, privacy-preserving browser-based web indexing approach that addresses privacy, autonomy, and offline access challenges; (2) defining theoretical adaphttps://doi.org/10.5281/zenodo.17229737 tive crawling and content acquisition methods based on semantic user-interest models that align content retrieval with preferences; (3) outlining efficient semantic embedding techniques optimized for browser-based storage and computation constraints; and (4) describing potential integration of browser-based language model capabilities supporting semantic search and retrieval-augmented generation for personalized content. RELATED WORK Browser technologies have advanced substantially from basic rendering to sophisticated local computation capabilities. Extending WebLLM’s work [Ruan et al . (2024)], researchers have further optimized on-device language model inference, reducing memory requirements and improving execution speed. These technical advances complement privacy-enhancing technologies research, where [Kumar et al . (2025)] developed frameworks for evaluating privacy preservation in AI applications without functionality compromises. Vector space representation of web content has enhanced information retrieval beyond keyword matching. Recent embedding techniques capture semantic relationships and contextual nuances that keyword approaches cannot address. Embedding compression methods have reduced storage requirements by up to 75% while maintaining 90% of semantic integrity [Li et al . (2024)]. These efficiency improvements are particularly valuable for browser environments with storage constraints. Research shows that optimized quantization and dimension reduction techniques maintain retrieval quality while reducing computational demands, balancing semantic precision with resource limitations. Adaptive crawling methodologies have proven effective beyond basic personalization. Building on [Durga et al . (2024)]’s user modeling work, subsequent studies have measured benefits showing up to 40% improvement in content relevance through dynamic crawling parameter adjustment. These approaches combine user interaction signals with content classification to create refined interest models. By analyzing content consumption patterns, dwell time, and explicit preferences, these systems develop accurate representations of user information needs that evolve over time. This adaptability particularly benefits specialized knowledge domains where standard crawling often misses relevant but less-connected content. Distributed indexing system architecture has evolved beyond simple centralized/decentralized divisions. The European OpenWebSearch.eu 1 project demonstrates how federated approaches can distribute computational workloads while maintaining consistent access patterns. Their federated storage approach separates crawling, indexing, and retrieval components, allowing specific optimization of each element [Granitzer et al . (2025)]. This architectural pattern informs our browser-based approach, though we adapt these principles to operate entirely within the client environment. 1https://openwebsearch.eu/ Content freshness maintenance in limited-resource environments represents another relevant research direction. Traditional search engines use continuous crawling with extensive server infrastructure, but resource-constrained systems require more strategic approaches. Recent research shows that selective recrawling based on content volatility prediction can maintain index freshness with reduced computational requirements [Gossen et al . (2015)]. These predictions use content type, historical update patterns, and domain characteristics to prioritize recrawling for rapidly changing content while conserving resources for stable information. While these research areas provide valuable foundations, integrating them into a cohesive browser-based indexing system presents unique challenges that remain insufficiently addressed. Current approaches tend to focus on individual components—either optimizing language models [Ruan et al . (2024)], improving vector representations [Li et al . (2024)], enhancing crawling strategies [Durga et al . (2024)], or developing distributed architectures [Hendriksen et al . (2024b)]—without fully considering how these elements interact within browser constraints. Our work synthesizes these advances into a comprehensive framework specifically designed for browser environments, addressing the technical limitations and privacy concerns inherent in centralized search systems. By combining adaptive crawling, efficient semantic indexing, and local retrieval augmentation, we propose a system that balances performance requirements with privacy preservation. The following sections detail our conceptual architecture and operational workflow, demonstrating how these components work together to enable personalized web indexing directly within the browser. CONCEPTUAL ARCHITECTURE This section outlines a conceptual approach to browserbased web indexing designed to enhance privacy and personalization. The methodology addresses constraints of centralized search systems through client-side processing, storage, and retrieval techniques that function within web browser limitations while enhancing user control. The methodology enables localized information management that reduces dependency on external search providers while maintaining search functionality. Content Acquisition The foundation of effective personalized indexing begins with selective content acquisition based on user interests. Unlike conventional web crawlers that aim for comprehensive coverage, this approach employs targeted crawling to retrieve only content aligned with individual user preferences, thereby reducing storage requirements while enhancing relevance. The system would construct dynamic user interest profiles through multiple mechanisms. Building on [Durga et al . (2024)]’s user modeling approach, the profile would incorporate both explicit inputs (user-specified topics, dohttps://doi.org/10.5281/zenodo.17229737 mains, and keywords) and implicit signals (browsing patterns, bookmarking behavior, and content interaction histories). These profiles would continuously evolve through adaptive algorithms that detect shifts in interests and adjust accordingly. Guided by these profiles, the crawling component would assign priority scores to potential URLs based on semantic alignment with user interests. This prioritization mechanism would consider both content similarity to established interests and exploration potential for adjacent topics. The crawler would maintain compliance with web standards and site policies, respecting robots.txt directives and implementing appropriate rate limiting to ensure responsible resource utilization. Beyond crawling methods, the system would offer alternative content acquisition pathways. Users can leverage the OWI Python client (owilix) developed by [Granitzer et al . (2025)], which provides sophisticated dataset management capabilities specifically designed for OWI environments. This tool enables efficient pushing and pulling of datasets and supports remote SQL query execution, allowing users to retrieve daily index slices precisely tailored to their interests without the overhead of full crawling operations. For users with private document collections, the system would implement a secure, privacy-preserving ingestion pipeline. This process begins with the secure parsing of personal documents stored in a self-hosted cloud solution, extracting valuable textual content and metadata. The extracted information is then normalized and loaded into DuckDB [Raasveldt and Mühleisen(2019)], a lightweight analytical database deployed within the user’s private infrastructure. This embedded database efficiently indexes the content, creating optimized structures for rapid querying. To enable seamless integration with client-side applications, the indexed content can be exported from DuckDB in JSON or similar serializable formats and imported into a compressed browser database. This final step bridges server-side indexing with client-side storage, providing users with efficient offline search capabilities while maintaining end-to-end privacy protection throughout the entire pipeline. Semantic Indexing Once content is acquired, the system would transform it into optimized representations suitable for browser-based storage and retrieval. The primary mechanism for this transformation would be dense vector embeddings that capture semantic relationships between content items beyond simple keyword matching. These embeddings would map textual content into multidimensional semantic spaces where proximity indicates conceptual similarity. Drawing inspiration from techniques described by [Li et al . (2024)], the system would generate embeddings at multiple granularity levels, from document-wide representations to sentence-level encodings. A key feature would be adjustable dimensionality, allowing dynamic balancing between semantic precision and storage efficiency. This adaptability would enable the system to operate effectively across devices with varying resource constraints. The processed content would reside in compressed browser databases utilizing technologies like IndexedDB [AlShaikh and Sleit(2017)]. To maximize storage efficiency within browser constraints, the system would implement structured data partitioning inspired by larger-scale approaches from the Open Web Index initiative [Granitzer et al . (2025)]. Content would be organized into logical segments based on source domains, temporal factors, and thematic categories, enabling efficient query processing. Additionally, metadata elements such as titles, content acquisition dates, and language indicators would be integrated directly alongside semantic representations to facilitate rapid filtering and result refinement during retrieval operations. Interactive Retrieval The retrieval process would begin with query encoding, transforming user information needs into the same semantic vector space used for content representation. These query embeddings would then undergo similarity comparison against the indexed content using established metrics such as cosine similarity, identifying the most relevant matches from the local database. Building on recent advances in browser-based AI frameworks demonstrated by WebLLM [Ruan et al . (2024)], the system would incorporate a locally executed language model for advanced retrieval and content synthesis. This model would implement retrieval-augmented generation (RAG) techniques, using the locally indexed content to ground its responses in user-specific information sources. The browser-native execution would leverage technologies like WebGPU [Kenwright(2022)] and WebAssembly [Haas et al . (2017)] to optimize performance within client-side constraints. User control would remain central to the retrieval process through customizable search parameters. These would include domain-specific weightings (prioritizing preferred sources), temporal filters (focusing on recent or historical content), and adjustable balance between semantic similarity and metadata matching. These customization options would allow users to tailor the system’s behavior to specific information-seeking contexts, from exploratory research to targeted fact-finding. Index Freshness Management Maintaining relevance over time requires mechanisms for content refresh and index optimization. The proposed system would implement context-aware scheduling for recrawling operations, prioritizing sources based on factors including update frequency, user engagement patterns, and content volatility. Instead of complete reindexing, the system would employ incremental processing techniques that efficiently integrate new content into existing indices. This approach would minimize computational overhead while ensuring the index remains current. The scheduling mechanism would balance https://doi.org/10.5281/zenodo.17229737 multiple factors: user preferences, connectivity conditions, and device resource availability, preferentially performing intensive operations during optimal conditions (e.g., during low-activity nighttime hours). Content pruning strategies would prevent unbounded index growth by identifying and removing outdated or lowrelevance items from the database. These decisions would consider multiple signals including recency, access frequency, and semantic redundancy with newer content. This comprehensive maintenance approach would ensure the system remains responsive and resource-efficient over extended usage periods while adapting to evolving user interests. OPERATIONAL WORKFLOW This section describes the conceptual workflow and component interactions in the proposed browser-based indexing approach. The design integrates various processes to enable personalized information access while maintaining user privacy and control throughout the operational cycle. Figure 1 shows an overview of the in-browser approach architecture and workflow. User Modeling Initialization The proposed system would begin with minimal setup requirements, avoiding intrusive information gathering during initialization. Instead of demanding extensive upfront configuration, the system would gradually build user interest profiles through two complementary mechanisms. The passive observation component would analyze content from past conversational search activities and pages visited during normal browsing in accordance with user privacy preferences. This lightweight semantic analysis would extract key concepts, entities, and topics without disrupting user experience. The extracted information would populate an initial interest model that evolves over time as the user continues browsing. Complementing passive observation, the system would provide explicit feedback mechanisms through which users could review, modify, or remove interests identified by the system. These controls would be prominently accessible within the browser extension, ensuring users maintain awareness and control over their interest profiles. Adaptive Crawling Strategy Once user interests are established, the content acquisition process would begin. The crawling component would employ a dynamic prioritization mechanism that evaluates potential URLs based on multiple factors: semantic alignment with identified interests, browsing frequency and historical engagement patterns. This prioritization would optimize resource allocation by focusing on content most likely to provide value to the specific user. To operate effectively within browser constraints, the crawler would implement adaptive resource management techniques. These would include adjustable parameters for concurrent requests, crawling depth, and scheduling frequency based on device capabilities and connection status. During active browsing sessions, the crawler would reduce its activity to minimize impact on performance, while potentially increasing activity during idle periods. The crawler would respect robots.txt directives, implement appropriate rate limiting, and follow standardized crawling policies. These practices would ensure the system behaves responsibly within the broader web ecosystem while gathering personalized content. In-Browser Indexing The indexing process would operate entirely within the browser environment, transforming retrieved content into searchable representations. Content processing would begin with semantic embedding generation, converting textual content into dense vector representations using locally stored or dynamically loaded models. These embeddings would capture semantic relationships between content items, enabling meaning-based rather than keyword-based retrieval. Following embedding generation, the system would extract and integrate metadata elements including titles, content acquisition dates, source information, and language indicators. This structured approach would enable efficient filtering during search operations. The indexed content would be organized using partitioning strategies inspired by the OWI project [Granitzer et al . (2025)], dividing the index logically by content origin, topical domains, or temporal factors. To maintain index freshness while minimizing computational demands, the system would implement incremental updating mechanisms. Rather than rebuilding the entire index when new content is acquired, only changes would be processed and integrated. A local changelog would track modifications enabling efficient updates. The system would also employ intelligent pruning algorithms to remove outdated or low-relevance content, preventing unbounded index growth over time. Retrieval-Augmented Search When users initiate a search query, the in-browser language model would process the input to understand the information need. The query would be encoded into the same vector space used for content representation, enabling direct comparison between the query and indexed content. The retrieval engine would identify relevant content based on semantic similarity measurements, returning results ranked by relevance to the user’s query. For complex information needs, the system would implement retrieval-augmented generation as described by [Ruan et al . (2024)]. This approach would ground language model outputs in the user’s personal index, combining the flexibility of generative AI with the accuracy of retrieved information. By leveraging locally stored content, responses would reflect the user’s specific knowledge base rather than generic information. The search interface would provide interactive refinement options, allowing users to adjust result presentation based on https://doi.org/10.5281/zenodo.17229737 Figure 1: An overview of the In-Browser indexing and personalized content retrieval approach. various parameters. These adjustments might include source preferences, recency requirements, or topic emphasis. Each interaction would feed back into the system’s understanding of user preferences, gradually improving retrieval accuracy through ongoing learning from user behavior patterns. User-Controlled Privacy Privacy protection would be fundamental to the system architecture, with all data processing occurring exclusively within the browser environment. This localized approach would ensure sensitive information remains under user control rather than being transmitted to external servers. The design would collect only information necessary for system functionality. The system would provide comprehensive transparency regarding data usage through an accessible browser extension interface. This interface would display the current interest model, crawling activities, and index contents in userfriendly formats. All aspects of the system would remain user-modifiable, with options to edit, export, or delete any stored information. Control granularity would extend to operational parameters, allowing users to adjust the balance between personalization depth and resource utilization. Users could configure crawling schedules, storage limitations, and embedding dimensions based on their preferences and device capabilities. This flexibility would enable the system to accommodate diverse usage patterns and hardware constraints while maintaining core functionality. Through this integrated operational flow, the proposed system would create a self-contained information ecosystem within the browser environment. By combining interest modeling, adaptive content acquisition, semantic indexing, and retrieval-augmented search, it would offer personalized information access while preserving user privacy. EVALUATION APPROACH Evaluating a browser-based indexing system presents specific challenges requiring careful methodological planning. This section outlines some possbile research-based approaches to assess such conceptual architectures. Technical performance evaluation requires adapting standard information retrieval metrics to the browser context. Measures such as precision, recall, and mean reciprocal rank must be applied within personal indexing constraints, where corpus size varies between users and changes over time. These metrics should assess retrieval effectiveness relative to indexed content rather than global repositories. Browser-specific indicators including memory usage, storage efficiency, and interface responsiveness are crucial for evaluating client-side feasibility. Simulation-based assessment offers valuable insights for conceptual architectures before full implementation. User simulation methods described by [Balog and Zhai(2025)] can be adapted to model various user interests, browsing patterns, and information needs. This enables systematic testing across different user profiles without extensive development resources. By creating synthetic browsing histories and interest profiles, researchers can generate representative personal indexes for testing. Simulated queries with predetermined relevance judgments provide measurable performance metrics while allowing parameter variation. LLM-based agents, following methods proposed by [Lu et al . (2025)], offer an effective evaluation strategy. These agents can simulate different user personas with varying information needs, technical expertise, and privacy concerns. This facilitates assessment of both technical performance and user experience aspects, including interface usability and perceived utility. While LLM agents cannot completely replicate human behavior, they provide cost-effective initial evaluation before human participant testing. Scientific validity requires careful benchmark development, including curated web content with predefined relevance judgments, standardized browsing profiles, and consistent query sets. Such benchmarks enable reproducible comparisons between implementation approaches and help assess design decisions regarding embedding dimensions, crawling strategies, and index partitioning methods. https://doi.org/10.5281/zenodo.17229737 Human participant validation remains essential for thorough evaluation. Well-designed user studies employing mixed methods can assess both objective performance and subjective experience. For this purpose, frameworks like SearchLab [Zerhoudi and Granitzer(2025)] offer valuable capabilities as a modular web-based platform specifically designed for search behavior studies. Participants should engage with the system over extended periods to capture realistic usage and allow natural interest profile development. Performance evaluation should combine logged interaction data and structured tasks with defined success criteria. Qualitative methods such as think-aloud protocols, interviews, and usability questionnaires complement quantitative measures by revealing user perceptions. The comprehensive data collection capabilities of SearchLab reduce the need for custom application development, allowing researchers to focus on study design rather than technical implementation. IMPACT AND RESEARCH DIRECTIONS The browser-based indexing approach we propose has implications beyond individual search experiences. This section examines potential effects on web ecosystems, user autonomy, and technological synergies, while outlining future research paths. Web Information Ecosystems Decentralizing web indexing through personal browserbased systems could alter web information dynamics. Current indexing power concentration among few corporations has created an environment where content discovery is mainly controlled by proprietary algorithms optimized for advertising revenue rather than information diversity. As [Granitzer et al . (2024)] note, this centralization introduces systematic biases that may homogenize content and favor commercial interests. A distributed approach where users maintain personal indexes could reduce these concentrating effects. Content creators might respond by producing more specialized material for niche audiences instead of optimizing solely for dominant search algorithms. Publishers currently invest in search engine optimization techniques that often prioritize algorithmic visibility over content quality. When discovery becomes more personalized through browser-based indexing, these incentives may shift toward content that serves user interests rather than algorithmic preferences. The proposed browser-based indexing system would function alongside broader open web initiatives. Users could optionally contribute anonymized, aggregated indexing data (with explicit consent) to collaborative projects like OpenWebSearch.eu [Granitzer et al . (2024)], creating a mutually beneficial relationship between personal and collective indexing efforts. This arrangement could address a limitation of purely personal indexing: reduced content discovery breadth. By voluntarily participating in federated efforts, users could maintain privacy advantages while contributing to and benefiting from collective knowledge organization. User Autonomy The architecture we propose improves user control over personal data and search experiences in several ways. By processing and storing data locally, the system removes the external data transfers found in centralized indexing systems. Users would gain protection from external data collection and clarity about what information their system has captured and how it affects their search results. The adaptive user-interest model offers another aspect of user empowerment. Unlike fixed indexing approaches that treat all users identically, the proposed system would refine its understanding of individual interests through browsing patterns and explicit feedback. This responsiveness allows search results to reflect actual user needs rather than general assumptions or commercial priorities. The system could show users visualizations of their interest profiles, allowing them to adjust or correct misinterpretations, enhancing both control and system accuracy. Clarity extends beyond data collection to the search process itself. Commercial search engines typically provide minimal insight into result selection for queries. A locally managed index could give users clear explanations of ranking factors, potentially building trust in the system. This clarity could help users develop better search strategies and understand the connection between their browsing behaviors and search outcomes. Leveraging AI Models Recent advancements in language models and generative AI create valuable opportunities for browser-based indexing systems. Local language models could improve multiple system aspects, from interest profiling to search query processing. By analyzing content semantics more deeply, these models could build more nuanced representations of user interests than conventional keyword-based approaches. This capability could help the system differentiate between temporary information needs and enduring interests, adjusting crawling priorities accordingly. The development and evaluation of such systems present distinct challenges that AI could help address. Language models could simulate various user behaviors to test system responsiveness across different usage patterns. While [Lu et al . (2025)] caution about limitations in AI-based simulation, such approaches could still provide useful insights during early development stages. These simulations could help identify weaknesses in crawling strategies or interest modeling before deployment with actual users. As browser-integrated language models like WebLLM become more capable, the system could implement proactive indexing based on anticipated information needs. The model might identify concepts related to current browsing activities and index relevant content in advance. However, such capabilities raise important questions about resource usage and user consent that would require careful consideration in any implementation. https://doi.org/10.5281/zenodo.17229737 Technical Challenges This proposal faces several implementation challenges. Browser memory and processing limitations represent a primary obstacle. Research is required to develop compact vector databases suitable for browser environments. Current embedding methods are typically designed for server environments with greater computational resources, requiring adaptation for client-side use. Techniques such as quantization [Li et al . (2024)] that reduce storage requirements while preserving semantic information could enhance the feasibility of the system. Adaptive interest models represent another research challenge. User modeling implementations must balance complexity with computational efficiency. Research into incremental model updates could improve user experience and resource use. Incorporating explicit feedback and implicit signals while maintaining model coherence presents a machine learning challenge requiring further study. As web content spans multiple modalities, research into efficient multimodal indexing becomes crucial. Extending browser-based systems to represent and search across text, images, audio, and video presents technical challenges. Unified embedding spaces that capture cross-modal relationships while remaining compact would advance the field. Evaluating personalized, decentralized search systems presents methodological challenges. Developing standardized benchmarks that accommodate individual differences while allowing systematic comparison would facilitate progress. Such frameworks need to address search quality, resource efficiency, and user satisfaction. These technical challenges highlight how browser-based indexing intersects information retrieval, machine learning, and human-computer interaction, requiring solutions that consider social and ethical implications of distributed information access. ACKNOWLEDGEMENTS This research was funded by the European Union’s Horizon Europe research and innovation program under grant agreement No 101070014 (OpenWebSearch.EU, https: //doi.org/10.3030/101070014). REFERENCES [Al-Shaikh and Sleit(2017)] Ala’a Al-Shaikh and Azzam Sleit. 2017. Evaluating IndexedDB performance on web browsers. In 2017 8th International Conference on Information Technology (ICIT). IEEE, 488–494. [Balog and Zhai(2025)] Krisztian Balog and ChengXiang Zhai. 2025. User Simulation in the Era of Generative AI: User Modeling, Synthetic Data Generation, and System Evaluation. arXiv preprint arXiv:2501.04410 (2025). [Durga et al.(2024)] Csl Vijaya Durga, RJ Anandhi, Saloni Bansal, Navdeep Singh, Ravi Kalra, and Nabaa M Bader. 2024. Adaptive Web Crawling Strategies Based on Ontological User Interest Modeling for Personalized Content Retrieval. In 2024 International Conference on Trends in Quantum Computing and Emerging Business Technologies. IEEE, 1–5. [Gossen et al.(2015)] Gerhard Gossen, Elena Demidova, and Thomas Risse. 2015. iCrawl: Improving the freshness of web collections by integrating social web and focused web crawling. In Proceedings of the 15th ACM/IEEE-CS Joint Conference on Digital Libraries. 75–84. [Granitzer et al.(2025)] Michael Granitzer, Mohamad Hayek, Sebastian Heineking, Gijs Hendriksen, Martin Golasowski, Michael Dinzinger, and Saber Zerhoudi. 2025. OpenWebSearch. eu-Building an Open Web Index on EuroHPC JU Infrastructures. Procedia Computer Science 255 (2025), 43– 52. [Granitzer et al.(2024)] Michael Granitzer, Stefan Voigt, Noor Afshan Fathima, Martin Golasowski, Christian Guetl, Tobias Hecking, Gijs Hendriksen, Djoerd Hiemstra, Jan Martinovič, Jelena Mitrović, et al . 2024. Impact and development of an Open Web Index for open web search. Journal of the Association for Information Science and Technology 75, 5 (2024), 512–520. [Haas et al.(2017)] Andreas Haas, Andreas Rossberg, Derek L Schuff, Ben L Titzer, Michael Holman, Dan Gohman, Luke Wagner, Alon Zakai, and JF Bastien. 2017. Bringing the web up to speed with WebAssembly. In Proceedings of the 38th ACM SIGPLAN conference on programming language design and implementation. 185–200. [Hendriksen et al.(2024a)] Gijs Hendriksen, Michael Dinzinger, Sheikh Mastura Farzana, Noor Afshan Fathima, Maik Fröbe, Sebastian Schmidt, Saber Zerhoudi, Michael Granitzer, Matthias Hagen, Djoerd Hiemstra, Martin Potthast, and Benno Stein. 2024a. The Open Web Index. In Advances in Information Retrieval, Nazli Goharian, Nicola Tonellotto, Yulan He, Aldo Lipani, Graham McDonald, Craig Macdonald, and Iadh Ounis (Eds.). Springer Nature Switzerland, Cham, 130–143. [Hendriksen et al.(2024b)] Gijs Hendriksen, Michael Dinzinger, Sheikh Mastura Farzana, Noor Afshan Fathima, Maik Fröbe, Sebastian Schmidt, Saber Zerhoudi, Michael Granitzer, Matthias Hagen, Djoerd Hiemstra, Martin Potthast, and Benno Stein. 2024b. The Open Web Index - Crawling and Indexing the Web for Public Use. In Advances in Information Retrieval - 46th European Conference on Information Retrieval, ECIR 2024, Glasgow, UK, March 24-28, 2024, Proceedings, Part V (Lecture Notes in Computer Science, Vol. 14612), Nazli Goharian, Nicola Tonellotto, Yulan He, Aldo Lipani, Graham McDonald, Craig Macdonald, and Iadh Ounis (Eds.). Springer, 130–143. https: //doi.org/10.1007/978-3-031-56069-9_10 [Kenwright(2022)] Benjamin Kenwright. 2022. Introduction to the webgpu api. In Acm siggraph 2022 courses. 1–184. [Kumar et al.(2025)] Priyanshu Kumar, Elaine Lau, Saranya Vijayakumar, Tu Trinh, Elaine T Chang, Vaughn Robinson, Shuyan Zhou, Matt Fredrikson, Sean M Hendryx, Summer Yue, et al . 2025. Aligned LLMs Are Not Aligned Browser Agents. In The Thirteenth International Conference on Learning Representations. [Li et al.(2024)] Xianming Li, Zongxi Li, Jing Li, Haoran Xie, and Qing Li. 2024. 2d matryoshka sentence embeddings. arXiv preprint arXiv:2402.14776 (2024). [Lu et al.(2025)] Yuxuan Lu, Bingsheng Yao, Hansu Gu, Jing Huang, Jessie Wang, Laurence Li, Jiri Gesi, Qi He, Toby https://doi.org/10.5281/zenodo.17229737 Jia-Jun Li, and Dakuo Wang. 2025. UXAgent: An LLM Agent-Based Usability Testing Framework for Web Design. arXiv preprint arXiv:2502.12561 (2025). [Raasveldt and Mühleisen(2019)] Mark Raasveldt and Hannes Mühleisen. 2019. Duckdb: an embeddable analytical database. In Proceedings of the 2019 international conference on management of data. 1981–1984. [Ruan et al.(2024)] Charlie F Ruan, Yucheng Qin, Xun Zhou, Ruihang Lai, Hongyi Jin, Yixin Dong, Bohan Hou, Meng-Shiun Yu, Yiyan Zhai, Sudeep Agarwal, et al . 2024. WebLLM: A High-Performance In-Browser LLM Inference Engine. arXiv preprint arXiv:2412.15803 (2024). [Zerhoudi and Granitzer(2025)] Saber Zerhoudi and Michael Granitzer. 2025. SearchLab: Exploring Conversational and Traditional Search Interfaces in Information Retrieval. In Proceedings of the 2025 ACM SIGIR Conference on Human Information Interaction and Retrieval (CHIIR ’25), March 24–28, 2025, Melbourne, VIC, Australia. ACM. https: //doi.org/10.1145/3698204.3716475 https://doi.org/10.5281/zenodo.17229737