Too much content, one screen
A large platform may have an enormous amount of content that could appear in your feed on any given day: posts from people you follow, videos from strangers, ads, suggested accounts. You can only look at a small fraction. Something has to decide the order, and that something is a recommendation system.
The companies involved generally keep the exact details private and change them often, so outsiders cannot describe any specific platform with certainty. The broad structure, however, is shared by most large services and is worth understanding.
The basic pipeline
Most feed systems work in stages, because scoring everything for everyone would be far too slow.
- Gathering candidates. The system narrows the full pool to a manageable set of items that might suit you, using things like who you follow, what similar users watched, and what is currently popular.
- Scoring. A model estimates, for each candidate, how likely you are to respond in various ways: watch to the end, click, like, comment, share, hide, or leave.
- Ranking. The scores are combined into a single ordering. Rules are then applied, such as limits on repeated authors, boosts for fresh items, and removal of content that breaks the platform’s policies.
- Showing and learning. What you actually do with the feed becomes new data, and the system updates.
The scoring step is a form of machine learning, trained on records of what past users did when shown past items.
What counts as a signal
The system learns from behaviour as well as from the content itself. Signals commonly include:
- how long you watched or read something, and whether you replayed it
- likes, comments, shares and saves
- what you clicked, skipped or scrolled past quickly
- what you explicitly hide, mute or mark as unwanted
- who you follow and interact with, and who is similar to you
- details about the item, such as its topic, age, and how others responded
- context, such as your device, language and time of day
Platforms differ in which signals count most, and in whether they treat a quiet pause as interest. Because behaviour is easier to measure than satisfaction, systems tend to rely on proxies for what you “want.”
Why engagement is the usual target
Many feeds are tuned to predict engagement, meaning actions such as time spent, clicks and shares. This is partly because engagement is measurable and partly because it connects to a platform’s business, since attention supports advertising, which is also why targeted advertising and feed ranking are so closely linked.
The problem is that engagement and benefit are not the same thing. Something can hold attention because it is useful and enjoyable, or because it is provocative, upsetting or compulsive. A system that only sees “the user kept watching” cannot always tell the difference. Many platforms say they also try to weigh surveys, quality measures and user controls, but how much those count varies and is hard to verify from outside.
What this means for what you see
Several side effects follow from this design, though their size is debated and varies by platform.
Feedback loops. If you watch a few videos on a topic, the system shows more, you watch more, and your feed narrows. That can be pleasant when the topic is baking, and less so when the topic is anxiety or conflict.
Popularity snowballs. Items that start well tend to be shown more widely, which can make a few pieces dominate while others never find an audience.
Strong emotion travels. Content that provokes a reaction may be shared more, and some researchers and commentators argue this can help misleading or inflammatory material spread, a concern often raised about online misinformation. How strong the effect is remains contested.
Uneven exposure. Creators often feel that reach changes unpredictably, because ranking rules are altered and rarely explained in full.
Your feed is not the whole picture. Two people can see quite different things from the same platform, so your feed is a poor guide to what “everyone” is seeing.
There are also links to wider worries, such as the effect of heavy use on social media mental health and on social media politics. Those topics are complicated, and a recommendation system is only one factor among many.
Recommendations beyond social media
The same ideas drive “you might also like” suggestions in shopping, music, news and video. Services such as video streaming platforms may rely more on what you have watched and what similar viewers chose than on what you follow. Shopping sites may add business considerations such as stock levels. The core is similar: predict what you will respond to, and show it.
Gaining more control
Without needing to understand any code, you can influence your feed.
- Use the feedback tools. Options like “not interested,” mute, block and hide are usually treated as strong signals.
- Choose your own ordering where offered. Some platforms offer a chronological view or a feed of only followed accounts.
- Be deliberate about what you linger on. Pausing on something counts for something, even if you disliked it.
- Review your settings and history. Some services let you pause or clear watch and search history, or switch off personalised suggestions. What is available depends on the service and region.
- Seek information on purpose. For news and important topics, visit sources directly instead of relying on a ranked feed.
- Take breaks. A fresh start sometimes loosens a narrow feed.
Recommendation systems are neither neutral mirrors nor sinister puppeteers. They are prediction tools aimed at goals chosen by companies, and knowing the goal helps you decide when to follow their suggestions and when to look elsewhere.