Auditing YouTube through TRACE

Auditing large-scale recommender systems like YouTube remainsa methodological challenge due to the trade-off between behav-ioral realism, scalability, and reproducibility. We present TRACE, anopen-source framework that integrates Large Language Models tosimulate context-aware, persona-driven user journeys. TRACE com-bines containerized browser automation, database-backed trace-ability, and asynchronous data enrichment to enable reproduciblelarge-scale audits of YouTube’s recommendation ecosystem. Bydecoupling experimental contexts from personas and supportingmultiple behavioral modes, it allows researchers to model diverseuser identities and explore how recommendation dynamics evolveover time. The framework’s modular architecture, reproducibledesign, and web-based control interface facilitate transparent, com-parative studies of algorithmic personalization and potential filterbubble formation. TRACE establishes a scalable foundation forempirically grounded, extensible, and ethically sound auditing ofrecommender systems.