It is Friday night, the snacks are ready, and you open Netflix with no idea what to watch. Twenty seconds later you are three minutes into a series you had never heard of.
That was no lucky guess. The Netflix recommendation algorithm built your home page for you, row by row and even poster by poster.
What is the Netflix algorithm?
“The Netflix algorithm” is really a family of algorithms. Together they decide which titles you see, in which rows, in what order, and with which artwork. Every member gets a different home page, and it changes as you watch.
Why go to all that trouble? Because people are impatient. Netflix’s own consumer research found that a typical member loses interest after perhaps 60 to 90 seconds of choosing, having looked at 10 to 20 titles. After that, the risk that they give up on the evening, or on the subscription, goes up.
And the audience is huge. Netflix passed 325 million paid memberships in the fourth quarter of 2025 (shareholder letter), which the company says is an audience approaching one billion people.
So how much rides on the recommendations? In 2015 two Netflix executives wrote that recommendations influence the choice for about 80% of hours streamed on Netflix, with the remaining 20% coming from search. They estimated that personalization and recommendations together save the company more than $1 billion a year, mostly by keeping members from cancelling (Gomez-Uribe and Hunt, 2015).
The story of Netflix recommendations: from DVDs by mail to AI
Netflix did not start out as a recommendation powerhouse. It started by mailing DVDs, and its first recommender had a very different job.
2006: a million dollars for ten percent
Back then, Netflix asked members to rate movies from one to five stars, and a system called Cinematch predicted how many stars you would give a movie you had not seen.
In 2006 Netflix did something unusual for a company: it opened the problem to the world. The Netflix Prize offered $1 million to anyone who could make Cinematch’s predictions 10% more accurate.
A year into the contest, a team called Korbell won the first progress prize with an 8.43% improvement, after more than 2,000 hours of work and a blend of 107 algorithms. Netflix took the two strongest ideas from that blend, matrix factorization and Restricted Boltzmann Machines, and put them into production.
And the grand prize? Someone did win the million two years later. The twist is that Netflix never put the winning solution into production: the extra accuracy did not justify the engineering effort. By then the company had a bigger thing on its mind.
2007 to 2012: streaming changes the question
Netflix launched streaming in 2007, a year after the prize began, and that changed everything about recommendations.
With DVDs, you picked a movie days before you watched it, so predicting your star rating made sense. With streaming, you want something good right now, you can try three titles in ten minutes, and Netflix can see whether you finished what you started. Viewing became far more useful than ratings.
By 2012 Netflix had more than 23 million subscribers in 47 countries, and 75% of what people watched came from some sort of recommendation. The team summed up its new philosophy in three words: “Everything is a recommendation.”
2015: the home page becomes the product
Next came the page itself. Netflix described a home page of about 40 rows, with up to 75 titles in each, where both the rows and their order are chosen for you. That is when the familiar rows took shape: Top Picks, Because You Watched, Continue Watching, Trending Now.
2017: thumbs up, and posters made for you
In 2017 Netflix swapped the five stars for a simple thumbs up or thumbs down. In tests, the change brought a 200% increase in ratings activity. People simply found it easier to say “liked it” than to decide between three and four stars.
The same year Netflix explained how it picks the artwork for each title per member. More on that fun one below.
2025: one big model and a new home page
In March 2025 Netflix described a foundation model for recommendations: one large model trained on hundreds of billions of interactions that many smaller recommendation features can draw on, instead of each team training its own from scratch.
Two months later Netflix unveiled a redesigned TV home page, with recommendations more responsive to your moods and interests in the moment, plus a small beta of generative AI search on iOS where you can type “I want something funny and upbeat”. For 2026 Netflix lists real-time personalized recommendations among the features it keeps working on.
From star ratings for DVDs to a model that reacts to your mood: twenty years, same goal. Help you find something to watch before you give up.
How does Netflix’s recommendation system work?
Here is the simple version. Netflix builds your home page in layers: it picks the rows, fills each row with titles, orders everything, and then chooses the artwork.
Rows, and why the top left matters most
Your eyes do most of the work here. Netflix found that members scan down the page more than across it, so titles in the upper left are much more likely to be seen than those in the lower right. That is why the best matches for you sit in the top rows and at the start of each row.
Each row comes from its own algorithm. The ones Netflix described in its 2015 paper:
- Personalized Video Ranker: orders the titles inside genre rows such as “Suspenseful movies” for you.
- Top-N Video Ranker: finds the few best picks from the whole catalog for the Top Picks row.
- Trending Now: catches short-term trends, from yearly ones like romantic movies around Valentine’s Day to one-off events like a hurricane in the news.
- Continue Watching: guesses which unfinished show you will actually go back to.
- Video-video similarity, or “sims”: powers the Because You Watched rows.
- Page generation: picks and orders the rows themselves, balancing relevance and variety, since one account often serves a whole household in different moods.
Collaborative and content-based filtering, Netflix style
Under those rows sit the two classic ideas behind most recommendation engines. Collaborative filtering looks for members who watch like you and suggests what they enjoyed. Content-based filtering looks at the titles themselves (genre, cast, release year, mood) and suggests similar ones. Netflix mixes both, which is why a brand-new show with no viewing history can still land on the right home pages. Our guide to what a recommendation engine is explains both with simple pictures.
What data does the Netflix recommendation system use?
Netflix lists the inputs on its help center:
- your interactions with the service, such as your viewing history and how you rated other titles,
- other members with similar tastes and preferences,
- information about the titles: genre, categories, actors, release year and so on,
- the time of day you watch, your language preferences, the devices you use and how long you watch.
And one thing it does not use: Netflix says the system does not include demographic information such as age or gender. Your taste is read from what you watch, not from who you are on paper.
Personalized artwork: why your poster differs from your friend’s
Ever noticed that a movie looks different on a friend’s account? That is on purpose.
Netflix often has several images for one title and chooses which one to show each member. Its engineers used Pulp Fiction to explain the idea: a member who watches a lot of Uma Thurman movies is likely to respond to artwork with Uma Thurman, while a John Travolta fan is more likely to click on artwork with John Travolta.
It sounds like a small thing. But when you only have about a minute and a half of someone’s attention, the picture is often what decides the click.
How Netflix’s recommendation system works today: one foundation model behind many rows
The rows and rankers above still exist, but since 2025 the brains behind them have been reorganized. This part gets a little more technical, so grab a coffee.
Why one big model replaced many small ones
For years every row had its own model: one for Continue Watching, one for Top Picks, and so on, each trained separately on much the same data. Netflix found that maintaining them all had become costly, and an improvement in one model was hard to carry over to the others. Most of them also looked only at a member’s recent history, because serving a longer one took too long.
The answer, borrowed from large language models, was a single foundation model that learns member preferences once and shares what it learns with every row.
Your viewing history, read like a sentence
The model treats each member’s history the way a language model treats text. Every interaction becomes a token: Netflix merges repeated actions on the same title into one token and keeps the important details, such as total watch time. Each token carries the context of the action (time, device, country, duration) and the title itself (its ID, genre, release country).
Training works like GPT’s: predict the next token, here the next title a member will engage with. With a few twists that make sense for TV rather than text:
- Not every event counts the same. A five-minute trailer play should not weigh as much as a two-hour movie.
- The model can predict several next steps at once, so it learns long-term taste instead of chasing the very next click.
- It also predicts side targets such as genre, which narrows down the candidates before it picks a title.
The scale is LLM-like too. More than 300 million members produce hundreds of billions of interactions, and Netflix reports that the usual scaling law holds: bigger models trained on more data keep getting better. One limit is very un-LLM-like, though. Recommendations must come back in milliseconds, so the model sees hundreds of events at serving time rather than a whole history; sparse attention and sliding windows during training help it learn from longer ones.
How new shows get recommended on day one
A brand-new series has no viewing history, which is the classic cold-start problem. The foundation model solves it by giving every title two descriptions: a learned ID embedding and an embedding built from metadata such as genres, storyline and tone. An attention layer mixes the two based on the title’s age, so a new release leans on its metadata and an established hit on its viewing data.
From the model to your home page
Netflix described the production setup in November 2025. The foundation model is pretrained from scratch every month and fine-tuned every day on the latest data, and the daily run also adds newly launched titles. Its learnings then reach the row models in three ways:
- Embeddings: compact profiles of members and titles, refreshed daily and published to an Embedding Store. Cheap and easy, the default starting point, but a day old by the time they are used. Netflix has since built a near-real-time version that updates during your session.
- Subgraph: the foundation model runs inside a row’s own model at request time, so nothing is stale. More powerful, but heavier, so it is kept for high-impact uses.
- Fine-tuning: a copy of the model tuned for one surface. For Trending Now, recent interactions with trending titles matter more than old ones.
Next up: language models as rankers
The newest step brings actual language models into the ranking. In a paper from August 2026, Netflix researchers describe GenRec, an open-source LLM adapted to Netflix data and then trained to rank titles, reading a member’s activity as plain text instead of hand-built features. In an A/B test it beat the production ranker with statistically significant gains, while using fewer labeled training examples. It is a research paper, so treat it as a direction rather than a description of every home page.
Is all this effort worth it? A study by Netflix researchers, first published in November 2025, estimated that replacing the current recommender with a plain popularity list would cut engagement by 12%. At Netflix’s scale, 12% of viewing is an enormous number of hours.
What Netflix’s approach means for your business
You don’t need Netflix’s scale to use the same ideas. Rank by what each person did last, put the best match top left, and test whether it works. We did exactly that with Recostream, our own recommendation engine for online stores: it returned recommendations in 20 to 30 milliseconds and stores measured a typical 5 to 10% sales uplift. GetResponse acquired it in 2022 (case study).
Running a streaming service? Our guide to the movie recommendation system covers the build for that case.
Recommendation engines in other streaming services
- Spotify: Discover Weekly, launched in 2015, has passed 100 billion tracks streamed and sparks more than 56 million new artist discoveries a week.
- YouTube: recommendations drive more viewing than subscriptions or search, and the system weighs clicks, watch time, surveys, shares, likes and dislikes (YouTube, 2021).
- TikTok: the For You feed counts strong signals, such as watching a longer video to the end, far more than weak ones (TikTok, 2020).
- Amazon Prime Video: runs on Amazon’s own engine; see how the Amazon recommendation system works.
Want recommendations like these in your own product? Our AI integration pilot adds them to an existing platform in about four weeks, with an A/B test to prove they work.