A recommendation platform sequentially collects information about a new product revealed from past consumer trials, and uses it to better guide later consumers. Because consumers do not internalize the value of information they bring to others, their incentive for trying out the product can be socially insufficient. Given such a challenge, I study how the platform can improve social welfare by designing its recommendation policy. In a model with binary product quality and general trial-generated signals, I find that the optimal design features a U-shaped sequence of recommendation standards over the product’s life, and the optimal learning dynamic can involve temporary suspensions following negative consumer feedback when the product is young. Comparative statics and extensions explore how the optimal design adjusts under changes in trial informativeness, consumer arrival rates, and platform bias. My analysis also illustrates the usefulness of a Lagrangian duality approach for dynamic information design.