Recommendation platforms (e.g., Google News, Netflix, or Amazon) often rely on social learning to improve their recommendations. Namely, they collect information generated from early consumer trials of a product (e.g., likes and reviews) and use it to better guide later consumers. In this process, however, because consumers do not internalize the informational value they generate for others, they often lack incentive to try a new product. This hampers learning and recommendation efficiency. To address this challenge, the paper studies how the platform should design its dynamic recommendation policy, which may “persuade” consumers toward more efficient social experimentation.

Unlike existing studies on this topic, the paper adopts a Lagrangian duality approach to handle consumers’ incentive compatibility constraints. This allows it to accommodate general non‑conclusive trial‑generated information, and enables characterization of the optimal design in terms of the recommendation standards evolving over time.

The paper finds that the optimal design features a U‑shaped sequence of recommendation standards over the product’s life. In particular, in the early phase, the standard should be gradually lowered as the product ages. A rough intuition is that when the product is very young, consumers are “skeptical” about following recommendations because they know the platform has not yet acquired much information. This necessitates a stringent selection regarding when to recommend the product. As time passes, consumers expect the platform to become better informed, making them easier to convince. The standard can hence be lowered. One implication of this result is that when the product is relatively young, it may be optimal to temporarily—not permanently—suspend recommendations following negative consumer feedback.

The paper also considers comparative statics and extensions exploring how the optimal design should adjust to changes in trial informativeness, consumer arrival rates, and platform bias. In particular, when trials become more informative or consumers arrive more frequently, the optimal recommendation standards should be lowered throughout the product’s life. The paper also highlights the general usefulness of the Lagrangian duality approach in dynamic information design problems.

About the Researcher 

Chen Lyu is an assistant professor at Peking University HSBC Business School. He holds a PhD in Economics from the University of Wisconsin-Madison. His research focuses on microeconomic theory, information design, digital economics, and financial market design.