Technology

How recommendation feeds decide what you see

By Leandro Bruzaferro · · 5 min read
A smartphone showing an endless scrolling feed
SharePostLinkedInFacebook

A recommendation feed feels like it knows you, and the explanations people reach for tend to be either mystical or paranoid. The actual mechanism is neither. It is a prediction engine trained on the behaviour of very large numbers of people, optimised for a target that is measurable, and the gap between that target and what you would say you want explains almost everything users find strange about it.

The signals that feed a ranking

Ranking systems consume three broad kinds of input.

Explicit signals are the ones you would name if asked: accounts you follow, things you liked, subscriptions, searches, an explicit “not interested”. They are trustworthy and scarce, because most people rarely use them.

Implicit signals are the behavioural exhaust of ordinary use: how long you paused on an item, whether you played it to the end, whether you replayed it, whether you opened comments, whether you scrolled past quickly. They are abundant, and they are what the system leans on, because there are thousands of them per session and only a handful of explicit ones.

Dwell time deserves particular attention, since it does most of the work. Slowing down is measurable, involuntary and available for every item you see. It is also ambiguous. You linger on things that fascinate you, and you linger on things that upset or confuse you, and the signal is identical.

Context signals cover device, time of day, connection quality and session length. These are less about you and more about which format fits the moment.

The last input is not about you at all: the aggregate behaviour of other users on the same item. A candidate that performs well with people who behave like you is promoted before you have interacted with it, which is why a feed can be accurate about someone new.

Engagement prediction, not preference

Here is the distinction that resolves most confusion. The system does not model what you like. It predicts what you will do.

A modern feed estimates several probabilities for every candidate: likelihood you watch a meaningful fraction, that you comment, that you share, that you follow the account, that you hide it or report it. Those estimates are combined into a single score using weights the platform chooses, and the weighting is the actual editorial policy of the product, expressed in arithmetic rather than in words.

Two things follow.

Content that provokes a reaction outperforms content you would rate highly on reflection, because provocation is easier to predict than satisfaction. Outrage, mild disbelief and unresolved curiosity all produce reliable behaviour. Quiet appreciation produces almost no measurable signal at all.

And negative feedback is a weak instrument. Hiding one item teaches the model very little, because a single negative sits against thousands of positive behavioural observations. This is why “show me less of this” feels ineffective. It is being counted, and it is being outvoted.

Platforms are aware of the mismatch and try to correct it, typically by adding survey-based targets that ask users whether they found something worthwhile and training against those answers. It helps at the margin. It does not change the underlying asymmetry, which is that behaviour is cheap to observe and judgment is expensive.

Why feeds converge on the same content

Recommendation systems have a structural bias towards items that already have evidence attached, since a candidate with a million observations can be scored confidently and a new one cannot. Exploration mechanisms exist to counter this, deliberately showing uncertain candidates to gather data, but the default gradient runs towards the already popular.

The result is convergence. Feeds across different platforms surface similar formats and similar subjects, not through coordination but because they are optimising similar targets on overlapping behaviour. It also produces the sensation of a feed narrowing over time: each confirmation makes the model more confident, and confidence reduces the variety it is willing to risk.

Whether this constitutes a filter bubble is genuinely contested in the research literature, and the honest position is that the effect is real but smaller and more uneven than the popular account suggests. What is not contested is the format convergence, which is visible to anyone.

Cold start and new accounts

A brand new account is the clearest window into how these systems work, because the personalisation has nothing to work with.

Early sessions serve broadly popular, low-risk content, sometimes weighted by location and language. Then the system watches. The first few dozen deliberate actions carry disproportionate weight, because they are the only evidence available, and they set the direction that subsequent behaviour refines.

This is why a fresh feed converges quickly and why it can converge somewhere unhelpful. Idle curiosity in the first hour is indistinguishable, to the model, from a stated interest.

What user controls actually change

Not all controls are equal, and knowing which are which is the practical payoff.

Control How much it does Why
Separate profiles or accounts per interest Most effective Prevents the linkage that personalisation depends on. Nothing else removes evidence this completely
Switch to chronological or following-only High The only setting that changes the objective rather than nudging the inputs
Clear or disable watch and search history High Removes the evidence the model is trained on, rather than arguing with the conclusion
Unfollow or mute a source High Removes candidates from consideration instead of adjusting a score
Per-item “not interested” Low One negative observation against thousands of implicit positives. It is counted, and outvoted
Topic preference toggles Low Usually influence the ranking rather than override it
Ad preference settings None, for the feed Govern advertising, not organic ranking. The interfaces rarely make this distinction clear

The pattern down that table is consistent and worth stating directly: the controls that work remove evidence, and the controls that disappoint try to argue with a conclusion the evidence already produced.

The useful way to think about all of this is that you are not stating preferences to a system that wants to satisfy them. You are producing evidence for a system predicting your next action. Behaving deliberately, especially early and especially with history settings, shapes the outcome considerably more than telling it what you like.

Sources