Why services recommend at all
A large streaming catalogue holds far more titles than anyone can browse. Without help, most viewers would scroll, give up or fall back on something familiar. Recommendation systems exist to shorten that search by putting a manageable set of likely choices in front of you.
They also serve the service’s own goals. A platform generally wants you to keep watching, keep subscribing and discover content it is promoting. The suggestions you see reflect a mix of what you might enjoy and what the service wants to surface, and the balance is not something outsiders can see.
The signals that feed suggestions
A system can only work with what it can observe. Although each service keeps its methods private, the following kinds of signals are commonly used.
- What you watched, and how much of it. Finishing a series says more than starting and abandoning it after five minutes.
- Explicit feedback. Thumbs up or down, star ratings, adding to a list, or removing a title from your history.
- Behaviour while browsing. Searches, what you hover over, and what you scroll past.
- Context. Time of day, the type of device, and sometimes the language or region.
- Details about the titles. Genre, cast, length, release year, and descriptive tags attached to each programme or song.
Not every service uses every signal, and the weight given to each is a design decision.
Two basic approaches
Most systems combine two broad ideas.
Content-based suggestions look at the attributes of things you liked and find similar titles. If you watch several slow-burning crime dramas, the system may offer more of that type. This works well with limited data, but it can keep you inside the same narrow style.
Collaborative filtering looks at patterns across many viewers. If people whose history resembles yours tended to enjoy a certain title, it may be suggested to you even if it looks nothing like what you have seen before. This can surface surprises, but it needs a lot of data from many users.
Modern services typically blend several methods, often using machine learning to predict how likely you are to click and how likely you are to enjoy a title once started. The training data behind such models is the pooled behaviour of many viewers, and any model can inherit the quirks of that data. Concerns such as algorithmic bias apply to recommendation systems too.
Profiles and shared accounts
Because suggestions are built from history, who is watching matters. If a household shares one profile, the history becomes a blend of everyone’s tastes, and the recommendations can look odd: a children’s cartoon beside a late-night thriller. Separate profiles keep each person’s signals apart. A guest watching on your profile can have the same distorting effect, which is why some services let you remove items from your history.
The cold-start problem
A brand-new account has no history, so the system has little to go on. This is often called the cold-start problem. Services usually respond by asking you to choose a few favourites when you sign up, or by showing what is popular in your region until your own habits emerge. New titles have a similar problem, since nobody has watched them yet. Descriptive tags and promotion placement often fill the gap.
Why recommendations can feel repetitive
Suggestions tend to narrow over time because the safest prediction is more of what you already watched. This is sometimes called a feedback loop: you watch what is shown, the system sees you watch it, and it shows more. Popular titles can also dominate because many viewers choose them, which reinforces their prominence.
It would be an overstatement to say these systems control what you like. They estimate and they steer, but taste is also shaped by friends, reviews, news and your own curiosity. The fair summary is that recommendations influence what is easy to find, not what you are able to choose.
Ways to steer what you see
You cannot read the algorithm, but you can give it clearer signals.
- Use the feedback buttons. A thumbs down or a request to see fewer similar titles is usually weighted more heavily than silence.
- Remove accidental viewing. If a service lets you delete an item from your history, do so for things you did not choose or did not like.
- Keep profiles separate for different people.
- Search deliberately. Looking up a genre, director or actor you want to explore tells the system something new, and often shows results outside the usual rows.
- Browse outside the home screen. Category lists, charts and curated collections are not personalised in the same way.
- Check privacy and personalisation settings. Some services let you pause history tracking or limit how data is used.
What this has to do with your data
Viewing history is personal data. It can reveal interests, routines and household makeup. Services generally keep it to improve their product, and some use it to inform advertising. If that concerns you, the privacy settings in your account, and general guidance on online privacy and how targeted advertising works, are worth a look. As with other data-driven systems, what a service collects and shares is described in its policy, which may change.