Netflix Rebuilt Its Recommendation Engine as a Language Model, and It Won the A/B Test by 0.115 Percent

Netflix put its long-running recommendation system head to head with a language model it built in house, and the language model won. The company published the results on its engineering blog on 22 August. The system is called GenRec, and its trick is almost disarmingly simple: instead of describing you with thousands of numbers, it describes you with sentences.

The old system relies on thousands of hand-crafted features, meaning small pieces of information that engineers designed and coded by hand, such as how often you finish a series or how much you like a given genre. That approach works, but it is expensive to extend. Every time Netflix adds a new content type like games, live formats or podcasts, someone has to build new features for it.

GenRec skips that step. Plays, watch durations, thumbs up or down, list additions and the moment you gave up on episode two all get written out as text, a bit like a running conversation between you and the service. A fine-tuned open-weight model, meaning a model whose weights are published so anyone can adapt them, reads that history and scores every matching title in one pass. Netflix does not say which open model it started from.

The numbers are modest but real. Offline, GenRec ranked about 1.6 percent better than the production system, and it got there with roughly 40 times fewer labelled training examples in its second training stage. In a four-week live A/B test on about ten percent of traffic, a short-term home-screen engagement metric rose 0.115 percent and a long-term core metric rose 0.006 percent. Netflix says both gains are statistically solid rather than noise.

What steers a recommendation now

Here is the part worth understanding, because it shows up everywhere in AI right now. The work has moved from feature engineering to context engineering. Nobody is designing clever numerical signals any more. The question is which events belong in the model’s input at all, and how many. A full text version of your viewing history would blow past the context window, meaning the amount of text a model can hold in mind at once, so Netflix filters hard: long sessions stay in full detail, quick taps get dropped, binges get condensed. Their own chart shows that past a certain length, each extra event barely helps and only costs compute.

Two honest caveats. Netflix calls this “an early but promising step” and is not replacing the old system. And these models go stale fast: a base model just two weeks old already loses ground because it has not seen the new titles.

What this means for you: If you use Netflix, nothing changes today, and a 0.115 percent shift is invisible to any single viewer. The interesting part is the pattern. A generic language model, fine-tuned on your own data and fed a well-chosen slice of context, is starting to beat systems that specialists spent years hand-tuning. If you run anything that ranks or sorts, that is the direction to watch. And if you are learning how AI works, remember the phrase: what you put in the prompt is now the design work.

Sources

• GenRec: Towards LLM-Native Recommendation at Netflix, Netflix Tech Blog
• Netflix tests language model as alternative to hand-built recommendation logic, The Decoder