Yu Song

Firms increasingly leverage personalization to influence product discovery and engagement throughout the customer journey. However, implementing effective multi-stage personalization is challenging, because the effectiveness of each stage depends on user intent and attention and cross-stage interactions can amplify or diminish overall impact. This paper examines personalization at two key stages: an early, firm-driven recommendation stage and a later, user-initiated search stage. Using data from a field experiment on a large e-commerce platform, I find that personalizing recommendations increases immediate revenue but decreases revenue in the later search stage, resulting in no net gain. I then develop and deploy a personalized search-ranking algorithm in a subsequent field experiment. The results show that search-stage personalization lowers search costs, enhances transactions, and does not cannibalize earlier recommendation-stage revenue. To explore applications of multi-stage personalization across different platform designs and heterogeneous consumers, I build a structural consumer search model. I exploit experimental variation and estimate the model using a neural network approach to address computational challenges. Counterfactual simulations reveal that personalizing only the search stage leads to higher revenue than personalizing both stages or only the recommendation stage. Moreover, personalizing personalization based on consumers' search costs further increases total revenue. These findings identify the conditions under which personalization is most effective, and offer guidance for firms on how to optimize their personalization across stages and consumers.

Content creators often find it difficult to anticipate and satisfy their audience’s needs, but the rise of online reviews offers a valuable source of real-time feedback. However, effectively incorporating feedback requires providers to invest time and leverage their experience. It remains unclear when and how producers should act on feedback during the development of a product series. If feedback reflects the majority point of view, making changes can increase popularity and signal care for customers. Nonetheless, if feedback only represents minority opinions, changes may not resonate with the broader audience. In addition, customers may lack the expertise to suggest changes that are feasible or optimal, and the suggested changes may not be compatible with the existing product. We examine when and how content producers respond to customer feedback shared through reviews, the consequences of those responses, and how feedback can be effectively used. Our study uses data from a major podcast platform. As episodes are released continuously, creators have repeated opportunities to respond to listener feedback. We first employ state-of-the-art natural language processing and large language models to extract actionable suggestions from unstructured review text, such as audio quality, content topic, voice‐to‐music balance, and speaking speed. We then apply deep learning methods and draw on behavioral and pedagogical theories to analyze audio data to examine whether those suggestions are implemented. Our results show that producers are most responsive to feedback delivered mid-series and on their most popular episodes. Moreover, early incorporation of listener input leads to significant increases in engagement in later episodes. We aim to identify the optimal timing and strategy for incorporating feedback. The findings will inform how platforms can solicit and present reviews, and guide content creators on how and when to incorporate feedback.

For companies with broad product portfolios, understanding consumer behavior across categories becomes essential for effective cross-selling. However, the mechanisms behind cross-category consumption remain underexplored, especially for non-complementary products. This paper asks two questions: When and why do consumers purchase products across distinct categories? How does a purchase in one category spill over into another? Leveraging individual-level data on grocery, insurance, financial-services, and vehicle purchases, we first document reduced-form cross-category patterns. For example, we find a complementarity between grocery promotions and credit-card ownership. A Causal Forest analysis shows that the promotional spillovers on purchase quantity and variety are strongest among younger, childless consumers. We then develop a structural model of consumer cross-category consumption decisions for both complementary and non-complementary products. Through counterfactual simulations, we aim to explore cross-category discount and bundling designs. Firms could use our findings to maximize the returns on their cross-selling efforts.

We investigate how platforms can balance the short-term and long-term effects of personalization. Analyzing data from a long-term personalized recommendation experiment on an e-commerce platform, we find that, in the short run, personalization reduces search friction and boosts platform revenue. However, in the long run, personalization narrows the variety of products displayed to consumers, leading to declines in engagement and retention. Sellers also strategically adjust their product assortments and pricing to compete for visibility in recommendations. These seller responses, in turn, impact consumer welfare and platform performance. We examine the long-run effects of personalization by accounting for both consumer responses and sellers’ strategic behavior. We aim to propose a personalization design that maximizes short-run conversion while promoting long-run consumer exploration and seller participation.