A social feed is not a list of recent posts but a ranked selection assembled fresh each time it loads. The stages of that assembly explain most of what people find strange about it.
Candidate generation comes before ranking
The system first gathers a pool of posts that could plausibly be shown, drawn from accounts followed, from groups joined and from material circulating beyond that circle.
That pool is far larger than a screen can display, and it already contains a great deal that a strictly chronological feed would never have reached.
Everything after this point is elimination, so a post that fails to enter the pool cannot be rescued by anything a viewer does later.
Ranking predicts actions, not preferences
Each candidate is scored by models estimating the probability of specific behaviours: a pause, a like, a comment, a share, a follow, a report.
Those probabilities are combined using weights the platform sets, and adjusting a weight changes what the whole population sees far more than any individual choice does.
The system is therefore optimising for predicted actions rather than for stated interests, which is why a feed can accurately reflect behaviour nobody would describe as a preference.
Engagement signals are not equal
A comment costs more effort than a like and is weighted more heavily, and content that reliably provokes replies is advantaged by that arithmetic.
Because disagreement produces comments efficiently, ranking systems tend to surface contentious material unless the weights are explicitly adjusted against it.
Platforms have added negative signals, such as hiding a post or marking it uninteresting, precisely to counterbalance what raw engagement would otherwise promote.
Filters run after the scoring
The ranked list passes through further passes that remove policy violations, limit how many posts from one account appear consecutively and inject unseen material.
Diversity rules exist because a purely score-ordered feed collapses quickly into repetition, which viewers experience as staleness and abandon.
The final order is therefore not the score order, and that gap accounts for many of the anomalies users notice and try to explain.
Feedback loops make the feed self-confirming
Every impression produces data, and that data trains the model that decides the next impression, so early signals compound into settled assumptions.
A brief interest can be reinforced into a dominant theme, while material that was never shown generates no evidence that it would have been welcome.
Deliberate use of the controls, following, muting and explicit feedback, is the only input strong enough to redirect a loop that otherwise runs on inference.