How Netflix Decides What to Recommend

Your Homepage Is Not a Catalogue, It Is a Prediction

When you open Netflix, you do not see a neutral list of everything available. You see a prediction: a homepage arranged around what the system thinks you are most likely to watch next, in what order, and with which images. That prediction is built from your past behaviour, from viewers with similar tastes, from information about each title, and from context such as time of day, device, and how long you usually watch. Netflix has said recommendations influence about 80 percent of hours streamed, with the rest coming from search. That figure, reported in Netflix’s own explanations and in coverage of its system, shows why the homepage matters more than the full catalogue for most viewers.

The homepage is also personal in a way that surprises first-time users. Two people in the same house, on the same account but different profiles, can see different rows, different title orders, and different artwork for the same show. That difference is not a bug; it is the system working as designed. Each profile has its own history, and the system uses that history to rank titles differently. A shared account with one profile, by contrast, mixes everyone’s tastes, which is why recommendations often feel confused in shared households. Separate profiles are the simplest way to keep predictions clean, and they are more important than any other homepage trick.

It helps to think of Netflix recommendations as three decisions, not one. First, which rows should appear, such as Trending Now, Because You Watched, or Top 10 in Your Country Today. Second, which titles should go in each row, and in what order. Third, which artwork, trailer, or preview should represent each title for you specifically. A title can be ranked highly but shown with the wrong image, and you may skip it. Or a title can be ranked modestly but shown with an image that matches your taste, and you may play it. Netflix optimises all three decisions together, because a recommendation that is not clicked is not useful, even if it was accurate in theory.

The Four Signals Netflix Says It Uses

Netflix’s own explanation lists four main signal groups. The first is your activity: what you watched, how long you watched it, whether you finished it, binged it, rewatched it, or abandoned it after ten minutes, and how you rated it with thumbs up, thumbs down, or Love This. Finishing a series quickly sends a stronger signal than starting it and stopping, because completion suggests satisfaction. Abandoning after ten minutes sends a different signal: the title may have been interesting enough to start, but not enough to continue. Rewatching sends the strongest loyalty signal, because it shows a title rewards repeat viewing.

The second signal is similar members: what people with tastes like yours have watched and enjoyed. This is collaborative filtering in plain language. If viewers who liked the same five shows as you also liked a sixth show, the system may recommend that sixth show to you, even if you have never searched for its genre. That logic allows Netflix to recommend titles you would not find by genre alone, because it uses human behaviour, not only labels. It also explains why recommendations can feel uncannily accurate: they are based on patterns across millions of viewers, not only on your own history.

The third signal is information about titles: genre, category, cast, director, release year, tone, themes, and other metadata. Netflix tags titles in detail, so it can connect shows that feel alike even when their plots differ. A slow, character-driven drama and a slow, character-driven thriller may share tone and pacing, even if their genres differ. Title information also helps with new titles that have no viewing history yet, because the system can match their tags to viewers who liked similar tags. Without title information, a new release would be invisible until many people had already watched it.

The fourth signal is context: time of day, device, and viewing habits. A viewer who watches short comedies on a phone during lunch may see different recommendations in the afternoon than in the evening on a television, when longer dramas are more likely. A new profile with no history sees more popular, broadly appealing titles, while a long-standing profile sees more personalised, niche titles. Netflix has said it does not use demographic details such as age or gender for recommendations, relying instead on behaviour, similar viewers, title information, and context. That claim matters, because it means your homepage reflects what you do, not who you are on paper.

From the Netflix Prize to a Foundation Model

Netflix’s recommendation history began with a famous public challenge. From 2006 to 2009, Netflix ran the Netflix Prize, offering one million dollars to anyone who could improve its rating predictions by at least ten percent. A team called BellKor’s Pragmatic Chaos won, but Netflix did not fully use the winning system, because its engineering cost outweighed its benefit at the time. That history is useful because it shows a trade-off that still shapes recommendations: a more accurate model is not always worth its complexity, if it is slow, expensive, or hard to maintain. Netflix must serve predictions in milliseconds, to hundreds of millions of viewers, on many devices.

For years after the Prize, Netflix ran hundreds of separate algorithms for different rows, tasks, and contexts. One model might rank a row, another might choose artwork, another might handle search, and engineers had to maintain all of them. In 2025, Netflix described a shift to a foundation model: a single large AI model, trained somewhat like a language model, that reads a member’s history of interactions as a sequence and powers the homepage, search, and recommendations together. That shift does not mean one simple formula; it means a unified model that can learn patterns across tasks, rather than many isolated models that each learn only one task.

The foundation-model idea also explains why recommendations now update more quickly. In 2025, Netflix redesigned its television homepage with larger cards and top navigation, and added responsive recommendations that update in real time based on what you just searched or watched. If you search for a Korean drama, your homepage may soon reflect that interest, even before you finish an episode. That responsiveness is useful, but it can also feel intrusive, because a single search can shift your rows. Understanding that responsiveness helps viewers interpret sudden homepage changes: they often reflect a recent search or a finished episode, not a permanent change in taste.

Why the Same Show Looks Different to Different People

Personalised artwork is one of Netflix’s most effective, and least visible, recommendation tools. Each title has many possible images, and Netflix chooses the image it thinks will appeal to you, based on your history. If you often watch romantic stories, you may see a show’s romantic couple on its thumbnail. If you often watch action, you may see the same show’s action scene. The show has not changed; the invitation has. That personalisation increases clicks, because people respond to images that match their taste, even when they do not consciously notice the difference.

That system also creates a risk: artwork can mislead. A drama with one action scene may look like an action series to an action fan, and a comedy with one sad scene may look like a drama to a drama fan. Viewers who play a title based only on artwork may feel disappointed, not because the recommendation was wrong, but because the image promised a different emphasis. The best defence is to read the synopsis, check the genre and episode length, and watch a trailer before committing to a series. Artwork is a suggestion, not a summary, and treating it as a summary leads to abandoned series and confused recommendations.

Trailers and previews work similarly. Netflix may show a different trailer cut, or auto-play a preview, depending on your history and device. A viewer who watches trailers fully sends a signal of interest, even without playing the title. A viewer who skips previews quickly sends a signal of disinterest. Those micro-signals, search, trailer views, and preview skips, accumulate into a profile that is more detailed than a simple list of watched titles. They also explain why a title you only searched for once can keep appearing: the system treats search as interest, even if you never played the title.

The Cold Start, Popularity Bias, and Filter Bubbles

Recommendation systems have known weaknesses, and Netflix’s system is no exception. The first weakness is the cold start problem: a new viewer with no history, or a new title with no viewers, is hard to recommend accurately. Netflix handles new viewers with popular, broadly appealing titles and with onboarding choices, and it handles new titles with metadata, trailers, and promotion in rows such as New Releases or Trending Now. Those solutions work, but they mean new viewers initially see a generic homepage, not a personal one. Personalisation improves only after you watch, rate, and search consistently.

The second weakness is popularity bias: the system tends to favour titles that are already popular, because popular titles have more data and more similar viewers. That bias makes it harder for niche, foreign-language, or older titles to break through, even when they would suit a particular viewer. Netflix tries to counter bias with diversity mechanisms, such as mixing genres, languages, and eras in rows, and with editorial rows that highlight specific collections. Viewers can also counter bias themselves, by searching deliberately for niche titles, rather than only choosing from the top rows. Search is a way to tell the system what you want, not only what it thinks you want.

The third weakness is the filter bubble: if the system only recommends what you have already watched, you may get stuck in one genre, one country, or one tone. A viewer who watches only crime dramas may see only crime dramas, even though they might enjoy a comedy or a documentary. Netflix counters bubbles with varied rows, but viewers must also break their own bubbles by trying one different title per week, rating it honestly, and letting the system learn from that experiment. A filter bubble is not only a system failure; it is also a habit, and habits can be changed by deliberate choice.

How to Train Your Netflix Homepage in Practice

Training Netflix starts with profiles. Create a separate profile for each regular viewer, and do not share a profile across very different tastes, such as a child’s cartoons and an adult’s thrillers. A mixed profile produces mixed recommendations, and no amount of rating can fully untangle it. If you already have a mixed profile, consider creating a new profile and rebuilding it with deliberate choices, rather than trying to fix years of mixed history. A clean profile learns faster than a confused one, and it costs nothing to create.

Ratings are the second tool, and they work best when used consistently. Use thumbs up for titles you enjoyed, thumbs down for titles you disliked, and Love This, the double thumbs up, for titles you loved. Rate finished titles, not only started ones, because completion plus rating gives a clearer signal than either alone. Do not rate titles you have not watched, to manipulate the system, because false ratings create false patterns. Honest ratings, even for only ten titles, can noticeably improve a new profile’s recommendations, because they give the system clear anchors for similar-viewer matching.

Watching behaviour is the third tool, and it is often more powerful than ratings, because it is harder to fake. Finish titles you enjoy, and stop titles you dislike early, rather than leaving them playing in the background. Background playing sends a false completion signal, telling the system you enjoyed a title you ignored. Similarly, do not leave a title paused for days in My List without deciding: My List is a signal of intent, but a list that grows to hundreds of titles tells the system little about priority. Keep My List short, remove titles you no longer want, and play the top three first. A short, active list trains better than a long, forgotten one.

Rows, Search, and the Top 10: What Each One Means

Rows have different purposes, and understanding them prevents misreading the homepage. Because You Watched rows are personal, based on your history and similar viewers. Trending Now rows reflect current popularity, often in your country, and are less personal. Top 10 rows reflect daily popularity, not personal taste, and they change quickly. New Releases rows reflect recency, not quality. My List rows reflect your saved intent, not the system’s prediction. A title in Top 10 is popular, not necessarily suited to you, while a title in Because You Watched is suited to you, even if it is not popular. Judge rows by their purpose, not by their position.

Search is the most direct way to override recommendations. When you search for a title, a genre, a director, or an actor, you tell the system what you want right now, regardless of its prediction. Search results also use personalisation, ranking titles you are more likely to enjoy higher, but they still show a broader set than the homepage. Use search when you know what you want, when you want a niche title, or when your homepage feels stuck. A single deliberate search can shift responsive recommendations, so search with intention, not idly, if you want your homepage to stay stable.

The Top 10 deserves special caution. A title can enter the Top 10 because many people started it, not because many people finished or loved it. A controversial, heavily marketed, or easily started title can trend briefly, then fade. Do not choose a series solely because it is in the Top 10, and do not avoid a series solely because it is not. Check its synopsis, its episode count, and a short review, and ask whether its mood suits your evening. The Top 10 measures attention, not satisfaction, and attention is a noisy signal for personal taste.

Privacy, Data, and What Netflix Does Not Use

Netflix recommendations rely on viewing data, which raises a fair privacy question: what does Netflix know, and what does it do with that knowledge? The system knows what you watched, searched, rated, and how you watched it, on which device and at what time, because that data is needed to rank titles. It does not, according to Netflix, use demographic details such as age or gender for recommendations. That distinction matters, but it does not mean viewing data is unimportant. Viewing data reveals taste, habits, and routines, and viewers should understand that a personalised homepage is built from that data, not from magic.

Viewers can control some of that data through profile settings, viewing activity, and ratings. You can view and delete titles from your viewing history in account settings, which can improve recommendations if a title was watched by someone else on your profile, or if you no longer want it to influence predictions. You can also adjust playback settings, such as auto-play previews, to reduce accidental signals. Those controls are not perfect, and deleting history does not erase all system learning instantly, but they give viewers agency. A recommendation system works best when viewers understand it is a tool they can shape, not a judge they must obey.

It is also worth noting what Netflix does not do, according to its own explanations. It does not recommend based on what you say you want in a survey, because stated preferences often differ from actual behaviour. A viewer who says they love documentaries but only watches comedies will see more comedies, because behaviour is more consistent than self-description. That principle, that actions speak louder than claims, is central to collaborative filtering. It also explains why recommendations can feel blunt: the system trusts what you did, not what you intended to do. Changing behaviour, not changing a survey answer, is how you change recommendations.

My Take:

My view is that Netflix recommendations are useful, but they should not replace choosing. A homepage that always gives you what you already like will eventually bore you, because it removes discovery, surprise, and effort. I prefer to use Netflix as a starting point, not as a decision-maker. I check Because You Watched for one option, I search for one option I chose myself, and I compare them. That comparison keeps me in control, and it prevents the strange feeling of scrolling for forty minutes while a system that knows my taste fails to offer anything I want. The system is good at prediction; it is bad at desire, because desire changes with mood, company, and day.

I also think separate profiles are non-negotiable, and I am surprised how many households ignore them. A shared profile is like a shared diary: everyone writes in it, no one can read their own story clearly, and recommendations become a compromise that pleases no one. Creating a profile takes two minutes and saves hours of confused scrolling. If you live with family, create profiles for adults, for teenagers, and for children, and protect children’s profiles with appropriate settings. That separation is not about privacy from family; it is about giving each viewer a homepage that actually reflects their taste, rather than a homepage that reflects everyone’s taste badly.

My strongest criticism is of personalised artwork, because it can mislead even careful viewers. I have started dramas that looked like thrillers, comedies that looked like romances, and documentaries that looked like dramas, all because artwork emphasised one scene over the whole. Artwork personalisation is clever marketing, but it is not honest summarising. I would prefer Netflix to show the same official poster to everyone, with a clear genre label, and let viewers decide. Personalisation should rank titles, not disguise them. When artwork disguises a title, it creates distrust, and distrust makes viewers ignore recommendations altogether, which defeats the system’s purpose.

I am more optimistic about the foundation-model shift, because a unified model can learn that a viewer’s taste is not a single genre, but a pattern across genres. A viewer who likes slow, character-driven stories may enjoy a slow drama, a slow thriller, and a slow documentary, even though those genres differ. A unified model can learn that pattern, while separate row models might miss it. That pattern-learning is the real promise of AI recommendations: not that they know what you want before you do, but that they notice connections you missed. When Netflix recommends a title from a genre you never tried, based on pacing or tone, that is a useful discovery, not a random guess.

Finally, I believe the best Netflix habit is a weekly experiment: one title outside your usual rows, watched for at least twenty minutes, rated honestly afterwards. That experiment breaks filter bubbles, teaches the system new information, and often leads to a surprising favourite. If the experiment fails, you have lost twenty minutes and gained a clear thumbs-down signal. If it succeeds, you have gained a new genre, a new country’s cinema, or a new director to follow. Recommendations should expand taste, not shrink it, and expansion requires deliberate effort from the viewer. Netflix can predict your past, but only you can choose your future viewing, and that choice is worth protecting.

I would also add a warning about My List as a graveyard. Many viewers add titles in a burst of enthusiasm, then never return, and the list grows to hundreds of titles that no longer reflect taste. A list that large is not a plan; it is procrastination in list form. I recommend a monthly My List clean-up: keep ten titles, remove the rest, and move removed titles to a note if you truly want them later. A short list forces a real choice, and a real choice is what recommendations are supposed to help you make. Without that discipline, even a perfect algorithm cannot help, because it is trying to predict from a list you never intended to follow.

Another specific habit is to watch the first episode of a series at a proper time, not late at night when tired. A first episode watched while half-asleep will be abandoned, and that abandonment will teach the system you disliked a title you never fairly tried. A series deserves a fair first episode: a television, not a phone, a quiet room, and enough time to finish without pausing every five minutes. If after a fair episode you still dislike it, stop confidently and rate it down. That honest signal is more useful than a tired, half-watched episode that confuses the system and wastes your evening.

I also think Netflix should show viewers why a title was recommended, in plain words, not only as Because You Watched a single title. A recommendation based on pacing, tone, and similar viewers is more informative than a recommendation based on one previous show. If Netflix explained, this drama shares slow pacing and ensemble cast with two shows you finished, viewers could judge whether that reasoning matches their taste. Transparency would also reduce distrust, because viewers would see recommendations as reasoning, not as manipulation. Until Netflix provides that detail, viewers should treat Because You Watched as a hint, not as an explanation, and check genre, tone, and reviews themselves.

Finally, I believe parents should use Netflix recommendations as a teaching tool for children, not as a babysitter. A child’s profile that only recommends cartoons can still include varied cartoons: different countries, different styles, different stories. Parents who watch one episode with a child, discuss what the child liked, and then search together for a similar but different title, teach the child how to choose, not only how to click. That skill, choosing deliberately in a world of infinite recommendations, will matter far beyond Netflix. A child who learns to choose will not be trapped by an algorithm, because they will know how to search, how to rate, and how to try something new on purpose.

One more practical point is to check recommendations after a guest uses your profile. A friend or family member who watches one series on your profile can shift your rows for weeks, because the system cannot tell a guest from you. If your homepage suddenly looks unfamiliar, check recent viewing activity, remove guest titles, and rate your own recent titles again. That quick clean-up restores your profile faster than waiting for the system to forget. It is also a good reason to offer guests their own profile, or a guest profile, before handing over the remote.

Frequently Asked Questions

Netflix ranks titles using your viewing activity, what similar viewers enjoyed, information about each title such as genre and cast, and context such as time of day and device. It then decides which rows to show, which titles go in each row, and which artwork to display for you.

Does Netflix use my age or gender to make recommendations?

Netflix has said it does not use demographic details such as age or gender for recommendations. It relies on viewing behaviour, similar members’ tastes, title information, and context, rather than on demographic categories.

Why does the same Netflix show have different artwork for different people?

Netflix personalises artwork, choosing among many images for the same title the one it thinks will appeal to you based on your history. A romance fan may see a couple, while an action fan may see an action scene from the same show, even though the show itself has not changed.

How can I train the Netflix algorithm to give better recommendations?

Use a separate profile, rate titles honestly with thumbs up, thumbs down, or Love This, finish titles you enjoy, stop titles you dislike early, keep My List short, and search deliberately for titles you want. Avoid background playing, because it sends a false completion signal.

Why does Netflix keep recommending popular titles I do not like?

Popularity bias means titles with more viewing data are easier for the system to recommend, and Top 10 rows reflect popularity, not personal taste. Use search and Because You Watched rows for personal picks, and try one different title per week to break a filter bubble.

What is the Netflix foundation model for recommendations?

In 2025, Netflix described replacing hundreds of separate row algorithms with a single foundation model that reads a member’s interaction history as a sequence and powers the homepage, search, and recommendations together. That unified model aims to learn patterns across tasks, rather than optimising each row in isolation.