E-commerce platforms have become central to modern economies, shaping how customers discover, evaluate, and purchase products. In these online platforms, decisions such as which products to display, how to structure assortments, and how to price products directly affect both profitability and customer satisfaction. At the core of these challenges are two fundamental problems: choice modeling, which captures the complexity of customer demand, and assortment optimization, which guides platforms in designing product offerings to optimize business objectives. In this dissertation, we introduce new models and algorithms that integrate customer choice behavior with assortment optimization to better capture actual operational challenges faced by online platforms. First, we study assortment optimization for online platforms that operate as market facilitators. These platforms, such as Instacart and DoorDash, do not have direct control over product inventory, and thus a product offered to a customer may have a significant probability of being out of stock by the time the customer’s order is fulfilled. To address this issue, platforms often ask customers to specify not only a preferred option, but also a replacement option in case the preferred option is unavailable at fulfillment. We develop a theoretical framework for this problem and design an approximation algorithm with provable guarantees for the resulting optimization problem. Using data from Instacart, we demonstrate that explicitly modeling replacement behavior can substantially improve expected revenue. Second, we study the mixture of two-stage Luce (MTSL) model, a framework that captures non-rational behavior. Under this model, the customer first uses a set of partial rankings to form a consideration set and then makes a purchase according to a Multinomial Logit model. Remarkably, we establish a connection between the MTSL model and ranking-based choice models: under certain conditions, the MTSL model admits a closed-form representation as a ranking-based structure, with negative “probabilities” assigned to certain rankings. We then study the assortment optimization problem under this model and show that it can be solved by a linear program when there is a single mixture component. When there are multiple mixture components, the problem becomes strongly NP-hard. Finally, our empirical analysis shows that the MTSL model delivers superior predictive performance relative to RUM benchmarks and remains competitive with other powerful machine learning–based models. Finally, we study assortment optimization problems in which platforms earn revenue from both product sales and customer clicks on sponsored products. The goal is to determine which products to offer and which products to sponsor to maximize the total expected revenue from sales and advertising. To study this problem, we propose the Click-Based Consideration Set model, a novel three-stage framework that captures customer clicking and purchasing behavior. We show that the resulting assortment optimization problem is NP-hard, and develop a fully polynomial-time approximation scheme. Using clickstream data from Instacart, we demonstrate that the proposed model improves predictive performance compared to existing models and jointly optimizing sales and advertising revenue leads to substantial revenue gains.