Feed ranking is frequently discussed as though someone is choosing what you see. What is actually happening is an optimisation process, and its outputs follow from what it was told to maximise.
The basic mechanism
A ranking system predicts, for each candidate item, how likely you are to interact with it.
Those predictions are combined into a score, and items are ordered by score.
The predictions come from models trained on historical behaviour — yours and that of people similar to you.
Which means the system is not selecting content by topic or by quality. It is selecting by predicted response.
What gets predicted
Platforms use multiple predicted actions, weighted.
Clicks, time spent, likes, comments, shares, and in some systems predicted negative responses used to demote items.
The weights are product decisions, adjusted continuously, and changing them changes what the whole platform looks like.
Public disclosures have described weight changes producing large shifts in what circulated, which demonstrates how much the outcome depends on these choices.
Why emotive content rises
Not because anyone chose to promote it.
Content producing strong reactions generates more of the measured actions, particularly comments and shares.
Which means optimising for engagement selects for content that provokes, without any intent to do so.
Research has found that content expressing hostility toward an out-group is shared substantially more than other political content, which is exactly the material an engagement objective will surface.
The feedback loop
What you engage with trains the model about you.
Which means engaging with something out of irritation is indistinguishable, to the system, from engaging with it out of interest.
Users frequently report that the feed shows them things they dislike, and the mechanism is straightforward — they interacted with similar things.
The practical implication is that scrolling past without interacting is a stronger signal than engaging to disagree.
The filter bubble question
More contested than the popular framing suggests.
Studies have found that social media users encounter more ideologically diverse content than people relying on offline sources, because their networks are larger and looser.
Which complicates the bubble account without dismissing it, since exposure to opposing views has in some studies increased rather than decreased polarisation.
The honest position is that the relationship between algorithmic curation and polarisation is genuinely disputed among researchers who have access to the data.
The publisher effect
Where the consequences are clearest.
News organisations built audiences dependent on platform distribution, then found that distribution changing without notice.
Which produced repeated business model crises as platforms deprioritised news, and several publishers have restructured entirely around it.
Platform decisions to reduce news distribution, taken for their own reasons, have measurably reduced traffic to journalism.
Chronological alternatives
Most platforms now offer a non-ranked option, generally not as the default.
Which is worth using occasionally simply to see what the difference is, since the contrast is informative about how much curation is occurring.
The reason it is not default is straightforward — ranked feeds produce more engagement, which is what the business runs on.
What actually helps
Unfollowing rather than muting, since muting still counts as a followed source in some systems.
Not engaging with things that annoy you, which is harder than it sounds.
Going directly to sources you value rather than waiting for them to appear.
And treating the absence of something from a feed as no evidence at all about its importance.
Recommendation against following
A structural shift in how feeds are assembled.
Feeds were originally built from accounts you followed, ordered somehow.
Several major platforms now populate largely from content you have not followed, selected by predicted interest.
Which removes the connection between who you chose to see and what you actually see, and it changes the platform from a network into a recommendation system.
The commercial reason is that it removes the constraint of your existing connections and allows the system to find whatever engages you.
Transparency and access
Independent research on these systems has been limited by data access.
Regulation in some jurisdictions now requires platforms to provide researchers with data and to explain recommendation systems in general terms.
Implementation has been contested, and several platforms have restricted or ended research access programmes during the same period.
Which means the evidence base on effects remains thinner than the importance of the question warrants, and much of what is known comes from internal documents released by whistleblowers rather than from systematic study.
What the controls actually do
Most platforms offer settings affecting what is shown — topic preferences, not-interested signals, sensitive content controls.
Their effect is generally modest, since they are inputs to a model rather than filters, and the model weighs them against everything else it knows.
Time and attention
Several platforms have shifted stated objectives toward time spent and away from raw engagement counts.
Which produces different content — longer video in particular — and is not obviously an improvement, since maximising time spent has its own consequences.
Some have introduced measures intended to reduce compulsive use, generally optional and lightly promoted.