Netflix recommends titles with a set of separate ranking algorithms, not a single model, and ranks both the titles in each row and the rows on the page. In a 2015 paper, two Netflix executives said the recommender influences about 80 percent of hours streamed, with search covering the rest. The same paper put the combined value of personalization and recommendations at more than USD 1 billion a year, which is a company estimate and not an audited figure.
Key takeaways
- A typical homepage has about 40 rows and up to 75 titles per row, and different algorithms fill different rows.
- Netflix states that recommendations draw on viewing history, ratings, similar members, title information, time of day, languages, devices, and viewing duration.
- Netflix states that the recommendation system does not use demographic information such as age or gender.
- The 80 percent and USD 1 billion figures come from a 2015 company paper, and no newer public equivalent was found.
- Algorithms are judged by randomized A/B tests that track member retention and medium-term engagement.
One homepage, several algorithms
The clearest primary description is a paper by Carlos Gomez-Uribe and Neil Hunt of Netflix, published in ACM Transactions on Management Information Systems in December 2015. It describes a homepage of about 40 rows, with up to 75 videos per row; the numbers vary by device. Each row typically comes from one algorithm. Genre rows such as Suspenseful Movies are ordered by a personalized video ranker, which sorts the whole catalog, or a genre subset, separately for each member profile. That ranker is used so widely that it has to work well for general ranking, which limits how personalized it can be, so the authors say it works better when personal signals are blended with a healthy dose of popularity.
Other rows have their own methods. The Top Picks row uses a ranker optimized only for the head of the list. Trending Now looks at short-term patterns, from minutes to days, such as a seasonal rise in romance viewing around Valentine's Day or interest in documentaries when a major storm is in the news. Continue Watching ranks recently viewed titles by the likelihood that a member will resume, rewatch, or has given up, using signals such as time since viewing and where playback stopped.
Rows, the page, and search
| Part of the page | What it does | Inputs named in the paper |
|---|---|---|
| Genre rows | Orders the catalog for each profile | Personal signals blended with popularity |
| Top Picks | Finds the best few titles for a member | Same family of signals, tuned for the top of the list |
| Trending Now | Surfaces short-term viewing trends | Recent viewing patterns, with some personalization |
| Continue Watching | Orders recently watched titles | Time since viewing, point of abandonment, device |
| Because You Watched | Lists titles similar to one the member watched | A title-to-title similarity list that is itself not personalized |
| Page generation | Chooses and orders the rows | Row relevance and variety across the page |
| Search | Matches queries and suggests alternatives | Play data, search data, and title metadata |
The paper notes that, before 2015, row positions followed a rule-based template shared by all members. A personalized algorithm now selects and orders rows from a large pool, and it can leave out a row type entirely for one homepage. The authors add that most members have different moods from session to session and that many accounts are shared by a household, which is one reason pages offer a varied selection of rows.
What Netflix says about the data it uses
Netflix's Help Center page on its recommendations system, checked on October 24, 2026, lists the inputs in plain language. They include a member's interactions with the service, such as viewing history and ratings, other members with similar tastes, and information about titles such as genre and release year. The page also names time of day, preferred languages, devices, and how long a member watched a title. It states that the recommendations system does not include demographic information such as age or gender in its decision making.
New profiles can pick a few liked titles to start; the page says later activity supersedes those choices, and recent activity counts more than older activity. Feedback from each visit, including what is started, finished, and rated, continually updates the algorithms. Beyond titles, the paper describes evidence selection, which chooses what to show alongside a recommendation, including which image supports it. A later Netflix engineering post on artwork personalization describes using contextual bandits, an online learning method, to choose images for each member.
How the numbers were produced
The authors say the main target of recommendation changes is member retention, but retention is so high that a 0.1 percent change is a meaningful improvement, so engagement is used as a correlated measure. Candidate algorithms are tested in randomized A/B tests in which members are assigned to different experiences. They report that over years of development, personalization and recommendations reduced monthly churn by several percentage points, and that the combined effect saves more than USD 1 billion a year.
These are the company's own estimates from a 2015 paper. Streaming catalogs, device mixes, and the methods themselves have changed since, and the paper does not publish its data. The figures show how Netflix reasoned about value at that time; they are not a current audit. Outside parties have not independently verified them in the sources reviewed here.
The Netflix statements on inputs and demographics are on its Help Center page on recommendations.
FAQ
Does Netflix use age or gender to recommend titles?
Netflix says its recommendations system does not use demographic information such as age or gender. Its stated inputs are activity, similar members, title details, time of day, language, device, and viewing duration.
Is the 80 percent figure still current?
It comes from a 2015 paper and refers to hours streamed influenced by the recommender, with the other 20 percent coming from search. No newer public figure was found.
Does everyone see the same row of titles?
No. The order of titles within a row, and the choice and order of rows, can differ by profile.
Can a viewer influence what is recommended?
Netflix says ratings, what a member starts and finishes, and initial title choices all feed the system, with recent activity weighing more than older activity.
Related reading
- AI and human-computer interaction
- Building an AI-based customer service chatbot
- Choosing AI apps for the iPhone
- AI and the student experience
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Gomez-Uribe CA, Hunt N, The Netflix Recommender System: Algorithms, Business Value, and Innovation, ACM Transactions on Management Information Systems 6(4), December 2015
- How Netflix's Recommendations System Works, Netflix Help Center, checked October 24, 2026
- Artwork Personalization at Netflix, Netflix Technology Blog