Cascaded generative AI lifts e-commerce cart adds by 2.7%

A new paper on arXiv presents a cascaded generative framework for e-commerce recommendations that tackles the rigidity of traditional personalized storefronts. The approach splits storefront construction into two tasks: 1) placement-level theme generation (e.g., deciding which products or categories to highlight per page section) and 2) constrained keyword generation per placement to power product retrieval. To handle latency and cost constraints, the authors use teacher-student fine-tuning, where a smaller student model is distilled from a larger teacher, achieving performance close to closed-weight LLMs while remaining deployable in production.

In online A/B experiments, the framework delivered an estimated +2.7% lift in cart adds per page view over a strong baseline. The system fuses generative output with traditional pointwise rankers, preserving existing infrastructure for stability. The authors also contribute AI-driven content evaluation and quality filtering to ensure safe, automated deployment. This work offers a practical blueprint for e-commerce platforms to inject semantic cohesion and dynamic merchandising into storefronts without a full overhaul.