GenPage is a generative AI system developed by Netflix to replace its traditional multi-stage recommendation pipeline by directly generating personalized user homepages. GenPage leverages user history and request context as a prompt to produce the entire page, resulting in improved user engagement and reduced serving latency.
Netflix's previous solution relied on "complex, multi-stage pipeline, with separate components for candidate generation and ranking". This process was repeated for each row and for each entity within a row, followed by a final stage responsible for page layout. In contrast, GenPage adopts a single-step approach that unifies all three levels, item selection, row construction, and layout generation, within one model:
Large language models have shown that a single generative model can perform diverse tasks just by generating a response to a prompt. Inspired by this prompt-response paradigm, we trained a single generative model to build the homepage by directly answering one question:
Given everything we know about this user and this request, what homepage should we generate to maximize user satisfaction?
Besides simplifying the workflow into a single step, GenPage outperforms the previous approach by enabling whole-page optimization. Through post-training reinforcement learning (RL), it can account for interactions not only within individual rows but also across rows and at the item level:
For example, a Continue Watching row near the top of the page may strongly satisfy a user’s immediate intent, but also reduce how much of the page they browse. Netflix also highlights that GenPage is more flexible in adapting to different content types and more easily extensible to new product experiences and layout variations.
Two key findings from production use are that enriching the prompt was more impactful than scaling model capacity and that post-training RL delivered an additional, unexpected benefit by increasing homepage diversity and customization. More specifically, Netflix engineers observed that while both scaling the model from 120M to 900M parameters and enhancing the prompt improved performance, prompt enrichment yielded the larger gain:
the gap between the two is striking. Scaling the model from 120M to 900M parameters reduces WBC loss by roughly 1.3%, whereas the cumulative effect of enriching the context is around 6.9%. In several cases, a single well-designed context addition delivers a larger improvement than the entire ~7.5× model-capacity scaling.
However, they note that context enrichment has diminishing returns: once the input context becomes saturated, scaling model capacity is likely to become the primary driver of further improvements. More broadly, they suggest that prompt enrichment may prove more effective than model scaling in other industry-scale personalization settings.
The benefits of GenPage were validated through A/B testing, which showed statistically significant improvements on the core user engagement metric. Another major advantage brought by GenPage is a 20% reduction in end-to-end serving latency, contrary to the common assumption that generative models are inherently slow.
There is much more in the original article than can be covered here, including details on how the system was designed, trained, and post-trained, as well as the trade-offs encountered and lessons learned. Be sure to read the full article for a comprehensive overview.