Netflix Replaces Multi-Stage Recommendation System With Single GenAI Model for Homepages

Netflix has introduced GenPage, a generative AI system designed to streamline how it constructs user homepages. Rather than relying on a complex series of specialized components, the new approach consolidates the entire process—from selecting content to arranging it on the page—into a single model that directly generates personalized layouts based on user behavior and request context.

A Simpler, More Unified Approach#

Netflix's previous architecture required multiple sequential stages: one component identified candidate content, another ranked those candidates, and a final stage determined page layout. This multi-step process repeated across each row and within each row's items, creating operational complexity. GenPage replaces this fragmented pipeline with a unified architecture that answers a single question: "Given everything we know about this user and this request, what homepage should we generate to maximize user satisfaction?"

Whole-Page Optimization Unlocks New Benefits#

The consolidated model enables Netflix to optimize the entire page as a coherent unit rather than treating individual rows in isolation. By leveraging reinforcement learning during post-training, the system can account for how content choices at different levels—between rows, within rows, and at the item level—interact with one another. For instance, positioning a "Continue Watching" row at the top might immediately satisfy a user's viewing intent but simultaneously reduce the likelihood they'll scroll further down the page. GenPage's unified approach captures these trade-offs across the whole experience.

Prompt Design Outpaces Raw Scaling#

Production results revealed a counterintuitive finding: enriching the model's input context—the information fed as the prompt—proved far more impactful than simply increasing the model's computational capacity. While scaling the model from 120 million to 900 million parameters improved performance by roughly 1.3%, enriching the contextual input delivered a cumulative 6.9% gain. As Netflix engineers noted, individual context improvements sometimes exceeded the benefit of a sevenfold increase in model size. However, diminishing returns emerge once the input becomes saturated; at that threshold, larger models become necessary for further progress.

Real-World Impact#