GenPage：Netflixにおけるエンドツーエンド生成的ホームページ構築に向けて GenPage: Towards End-to-End Generative Homepage Construction at Netflix

Netflix has outlined a new system called GenPage that uses generative artificial intelligence to assemble a member's entire homepage through a single, end-to-end process rather than stitching it together from independently optimized parts. The work matters because the homepage is the primary surface through which the service's subscribers decide what to watch, and even modest improvements in how content is surfaced can meaningfully affect engagement and discovery across a very large catalog.

The Netflix homepage is typically structured as a vertical stack of rows, each with a theme such as "Trending Now" or "Because You Watched," with titles arranged horizontally within each row. For years, streaming and e-commerce platforms have constructed such pages using a multi-stage pipeline. Candidate generation first narrows an enormous catalog to a manageable set of items, ranking models then score those items for predicted relevance, and a separate page-construction stage decides which rows to include and how to order them. This modular design is efficient and easy to reason about, but because each stage is often optimized against its own local objective, the assembled result can be globally suboptimal. Common symptoms include the same title appearing across multiple rows, thematic redundancy between adjacent rows, or a layout that makes sense component by component yet fails to present a coherent whole.

GenPage, as described on the Netflix engineering blog, appears to address these limitations by treating the homepage as a single object to be generated rather than a collection of separately ranked pieces. Instead of ranking rows and items in isolation and then merging them, a unified model reasons about the page holistically, taking into account how rows and titles interact, complement one another, and collectively serve a member's tastes. The stated goal is to move beyond partial or per-component optimization toward integrated personalization, so that decisions about what to show and where to place it are made with awareness of the full context of the page.

Technically, this kind of approach is consistent with a broader shift in recommendation systems toward generative and sequence-based modeling. Page construction can be framed as a generation problem in which a model produces an ordered arrangement of rows and items, somewhat analogous to how a language model generates a sequence of tokens. Architectures built on transformers and attention mechanisms are well suited to capturing dependencies across a layout, allowing the system to consider diversity, novelty, and coherence jointly rather than after the fact. Netflix has not, in this summary, disclosed the full details of the model architecture, training objectives, or serving infrastructure, and any specifics beyond the general framing should be treated cautiously until the complete technical description is available.

The effort fits within a wider industry movement. Netflix has long been associated with recommendation research, dating back to the Netflix Prize competition that helped popularize collaborative filtering, and it has since published extensively on ranking, candidate generation, and artwork personalization. More recently, companies including large e-commerce and short-video platforms have explored generative retrieval and end-to-end learned ranking, reflecting a general interest in replacing hand-tuned, multi-stage pipelines with models that optimize an overall objective directly. GenPage can be read as an application of that philosophy to the specific challenge of composing a two-dimensional page layout, which is a harder problem than ranking a single list because it involves both vertical row selection and horizontal item ordering simultaneously.

Several practical considerations are likely to shape how such a system performs in production. End-to-end generation must still meet strict latency requirements, since homepages are rendered on demand across many device types, and it must remain controllable enough to respect business rules, editorial constraints, and content licensing. Evaluating a whole-page model is also more complex than measuring the accuracy of a single ranked list, typically requiring online experimentation to confirm that holistic construction genuinely improves member outcomes. If GenPage delivers on its stated aims, it would represent a notable example of generative AI being applied not to create media content itself, but to organize and present existing content more effectively, an area of personalization that tends to receive less public attention than generative text or imagery.