User Agents Transform Online Recommendation Markets

Traditional online recommendation systems typically operate within a single platform, where the platform controls the pool of candidates and their ranking. However, the emergence of large language model-based user agents is fundamentally altering this dynamic. These agents allow users to articulate their needs before selecting a platform, leading to a new paradigm called "agentic recommendation markets," where platforms actively compete for the user's attention. Experiments conducted across three product domains reveal a significant tension in this new environment: while user-centric recommendation vastly expands the potential pool of relevant items, this broader access doesn't automatically translate into effective user exposure. Platforms engage in strategic behaviors, such as using selectively positive explanations, which dominate first-ranked positions in the absence of user feedback mechanisms. Crucially, when user agents incorporate feedback on platform actions and subsequent user outcomes, the prevalence of these strategic explanations decreases, and the likelihood of a user purchasing the relevant item increases. This highlights that a user agent is more than just an advanced ranker; its design, including how it queries, ranks, and manages feedback, directly influences user utility and necessitates a joint mechanism design approach considering access, attention, and accountability.