How recommendation systems decide what you see

Open almost any app today and something has already decided, before you asked, what you should see next: a video, a product, a song, a post from someone you have not spoken to in years. These decisions are made by recommendation systems, and while the specific engineering varies between companies, most rely on a handful of well-established approaches that are worth understanding.

Collaborative filtering: learning from people like you

One of the oldest and most widely used approaches is collaborative filtering, which works on a simple premise: if two people have liked similar things in the past, they are likely to agree on new things too. The system does not need to understand what a piece of content is actually about; it only needs to know patterns of agreement across many users. This is why you might get a strong recommendation for something that seems to have nothing in common with your usual interests, purely because people with a similar overall pattern of behaviour to yours also liked it.

Content-based filtering: matching attributes

A different approach looks at the actual attributes of content rather than the behaviour of other users. If you have watched several documentaries about a particular topic, a content-based system will recommend more documentaries with similar attributes, such as genre, length, or subject tags. This method works well even for new content with little user history, but it can also create a narrower range of suggestions, since it tends to recommend more of the same rather than something genuinely different.

Combining approaches

Most large-scale systems in practice combine both methods, along with additional signals like the time of day, your device, recent trends, and how long you spent engaging with something rather than just whether you clicked on it. This hybrid approach tries to balance relevance, meaning content you are likely to enjoy, with discovery, meaning content that broadens what you see rather than narrowing it further.

What signals actually get used

• Explicit signals, such as a rating, a like, or a follow
• Implicit signals, such as how long you watched or read something before moving on
• Contextual signals, such as time of day, device type, or location
• Social signals, such as content shared or engaged with by people you are connected to

Implicit signals tend to carry more weight than most people expect, since they reflect actual behaviour rather than stated preference, and behaviour is generally a more reliable predictor of future interest.

Why this can create narrow feeds

A well-known side effect of recommendation systems is that they can gradually narrow what you see, since the system is optimising for engagement based on what has worked before, not for balance or breadth. This does not require any deliberate intent to mislead; it is simply what happens when a system is rewarded for keeping your attention and learns that certain types of content reliably do so.

How this connects to tracking

Recommendation systems generally rely on data collected about your behaviour, which connects closely to broader questions about cookies and tracking discussed elsewhere in our internet coverage. The more data a system has about your behaviour across contexts, the more precisely it can tailor recommendations, which is part of why cross-site tracking has been commercially valuable to companies that also run recommendation-driven products.

What you can actually influence

Most platforms with recommendation systems offer some way to shape what you see, whether by explicitly disliking content, adjusting settings, or clearing your watch or search history. These controls are often underused, partly because they are not prominently placed, but they can meaningfully shift what the system shows you if you are deliberate about using them.

A useful habit

Understanding that a feed is a product of specific mechanical choices, rather than a neutral reflection of what is popular or important, is itself useful. It makes it easier to recognise when a feed feels repetitive or narrow, and to take the small available actions that push it back toward something more varied.