The short answer
Your feed is not showing you what you asked for. It is showing you what a prediction system expects you to engage with. Here is how that machinery actually works.
- What you stopped on. Dwell time is one of the most informative signals available, and you produce it without any deliberate action.
- What you finished. Watching to the end, or rewatching, counts far more than a passing view.
- What you did next. Following a link, saving something, sharing it privately.
- What you avoided. Scrolling past fast, hiding, muting. Negative signals shape the feed as much as positive ones.
- Who you resemble. The behaviour of people with similar patterns predicts a surprising amount about you.
A recommendation algorithm does not decide what you should see. It predicts what you are likely to react to, then sorts everything available by that prediction. The feed you get is a ranked list of guesses about your behaviour, assembled fresh each time you open the app.
Once you see it that way, a lot of confusing experiences make sense. Why one offhand tap changes your feed for a week. Why the same account shows you completely different things than it shows a friend. Why the content you say you want and the content you actually get keep diverging.
The basic loop
Nearly every recommendation system, whatever the app, runs the same three-step loop.
First it gathers candidates. Out of everything that exists on the platform, a fast and rough process pulls a manageable pool of things you might plausibly want. This stage is about speed, not judgement.
Then it scores them. A model estimates, for each candidate, how likely you are to perform various actions. Watch it to the end. Like it. Comment. Save it. Close the app immediately.
Finally it ranks and adjusts. The scores get combined into one number, then modified by other rules: do not show the same creator four times in a row, mix in something new, respect content policies. What survives is your feed.
What it actually knows about you
The picture is built almost entirely from behaviour rather than from anything you declared.
- What you stopped on. Dwell time is one of the most informative signals available, and you produce it without any deliberate action.
- What you finished. Watching to the end, or rewatching, counts far more than a passing view.
- What you did next. Following a link, saving something, sharing it privately.
- What you avoided. Scrolling past fast, hiding, muting. Negative signals shape the feed as much as positive ones.
- Who you resemble. The behaviour of people with similar patterns predicts a surprising amount about you.
- Context. Time of day, device, connection speed, session length. Your morning self and your late-night self are treated as different audiences.
Notice that stated preference barely appears. Following an account is a weak, one-time signal. Watching a video twice is a strong, repeated one. The system believes what you do, and what you do is not always what you would say you want.
Two ways of guessing, working together
Under the hood, most recommenders blend two complementary approaches.
Collaborative filtering works from patterns across people. If many users who behaved like you enjoyed something you have not seen, it becomes a good candidate. The system does not need to know anything about the content itself, only who engaged with what. This is why recommendations sometimes cross topics in ways that feel uncanny.
Content-based filtering works from the item. What is this video about, what does it sound like, what text appears in it, which topics does it resemble. This is what allows brand new content to be recommended before anyone has interacted with it.
Modern systems represent both users and items as long lists of numbers, sometimes called embeddings, placed in a shared mathematical space. Things that behave similarly end up near each other. Recommendation becomes a search for nearby points. Nobody hand-labels the categories, which is why the resulting clusters can be oddly specific and hard to name.
What the system is optimising for
Here is the part that explains most of the friction. The algorithm optimises a business objective, and that objective is usually engagement, because engagement is measurable, immediate, and connected to revenue.
Satisfaction is much harder to measure. Whether something was worth your time, whether you felt better afterwards, whether you learned anything. These do not produce a clean signal within seconds, so they get approximated at best.
The gap between those two things produces predictable side effects. Content that provokes a quick reaction outperforms content that rewards patience. Outrage travels well because it reliably generates responses. Mild, useful, slow material struggles, not because anyone decided it should, but because it scores poorly on the proxy being measured.
Platforms are aware of this and do try to correct for it, with surveys, downranking of clickbait, and objectives that include longer-term retention. Whether those corrections are adequate is a genuine open argument, and reasonable people who study this disagree about how much of the problem is fixable while engagement remains the core measure.
Why your feed narrows
Recommenders face a constant tension between showing you more of what works and exploring something new. Too much of the first and the feed becomes claustrophobic. Too much of the second and it feels random.
The tilt is usually toward exploitation, because it performs better in the short run. That produces the familiar experience of a feed that keeps tightening around a topic until you have to actively break out of it.
The filter bubble question is more nuanced than the popular version suggests. Research findings are mixed, and some studies find people encounter more varied views online than offline. What is less disputed is that these systems amplify whatever tendency you already have. If you lean toward one kind of content, the feed will happily take you further in that direction.
How to actually steer it
You have more influence than it feels like, but only if you use the signals the system weighs heavily.
- Use the explicit controls. Not interested, hide, mute, do not recommend this channel. These are strong, direct signals and most people never touch them.
- Change your dwell time. Scroll past quickly when something is baiting you. Lingering to be annoyed reads as interest.
- Finish what you value. Completion is heavily weighted. Half-watching good content teaches the system you did not like it.
- Search deliberately. Explicit searches are a strong statement of intent and usually reshape recommendations faster than passive browsing.
- Separate your contexts. Different accounts or profiles for work and leisure stop one interest from contaminating the other.
- Clear old history. Where the platform allows it, pruning watch history removes the evidence keeping you in a rut.
A fairer way to think about it
It helps to drop the idea that the algorithm is a single entity with intentions. It is a prediction system tuned to a target somebody chose, running on evidence you generate continuously and mostly unconsciously.
That reframing is useful in both directions. It stops you assuming malice where there is only optimisation, and it stops you assuming the feed reflects your considered preferences. It reflects your behaviour, which is a different and less flattering thing. Adjust the behaviour and the feed follows.
Comments
0 total