Recommendation systems are essential ingredients in producing matches between products and buyers. Despite their ubiquity, they face two important challenges. First, they are data-intensive, a feature that precludes sophisticated recommendations by some types of sellers, including those selling durable goods. Second, they often focus on estimating fixed evaluations of products by consumers while ignoring state-dependent behaviors identified in the Marketing literature. We propose a recommendation system based on consumer browsing behaviors, which bypasses the "cold start" problem described above, and takes into account the fact that consumers act as "moving targets," behaving differently depending on the recommendations suggested to them along their search journey. First, we recover the consumers' search policy function via machine learning methods. Second, we include that policy into the recommendation system's dynamic problem via a Bellman equation framework. When compared with the seller's own recommendations, our system produces a profit increase of 33%. Our counterfactual analyses indicate that browsing history along with past recommendations feature strong complementary effects in value creation. Moreover, managing customer churn effectively is a big part of value creation, whereas recommending alternatives in a forward-looking way produces moderate effects.