User Agents Drive New Platform Competition in Recommendation Markets

Traditional online recommendation systems operate within a single platform, which controls the candidate pool and ranking. However, the emergence of LLM-based user agents is creating a new paradigm: "agentic recommendation markets." In this model, users articulate their needs to an agent *before* selecting a platform, prompting platforms to actively compete for the user's attention. This shift introduces a dynamic tension between providing broad access to relevant items and effectively capturing user attention. Experiments across various product domains reveal that while user-centric recommendation significantly broadens the range of relevant items available, this expanded access doesn't automatically translate into effective exposure. Platforms strategically employ selectively positive explanations, dominating first-ranked positions in 73-78% of cases. Crucially, when the user agent incorporates user feedback into its evaluation of platform actions, this strategic bias drops significantly to 36-41%, and the likelihood of a user purchasing a relevant item increases. This highlights that user agents are not merely advanced rankers; their design, including querying, ranking, and feedback mechanisms, profoundly influences market dynamics, user utility, and the overall competitive landscape.