How Recommendation Algorithms Work: Why Your Feed Shows What It Shows

Open any feed and it feels personal — sometimes uncannily so. The video you mentioned yesterday, the product you glanced at once, the creator you watched for eleven seconds last month: all of it returns, arranged. None of it is magic and little of it is spying; it is prediction, run at industrial scale. Understanding how recommendation algorithms work demystifies the feed: what signals the systems collect, how they rank a universe of content into your twenty items, and why the loop they create tends to narrow what you see. This explainer follows one recommendation from raw data to your screen.

The Raw Material: Signals You Generate Constantly

Recommenders eat behaviour. Every platform logs explicit signals (likes, follows, ratings, saves, searches) and far more implicit ones: what you watched and for how long, what you scrolled past and how fast, what you rewatched, shared, commented on, bought, or abandoned mid-way. Dwell time — the seconds your eyes rest on an item — is among the most honest signals ever collected, because you cannot fake attention you did not give. Context rides along: device, time of day, location, language, and the behaviour of people whose patterns resemble yours. The first thing to understand is asymmetry of effort: you produced this data as exhaust while doing other things; the platform refined it deliberately. If the vocabulary here feels new, this site’s primer on machine learning basics covers the underlying ideas — models learning patterns from examples — in beginner terms; recommendation is that machinery pointed at your attention.

The Two Stages: Retrieval, Then Ranking

No system scores every video or post in existence for every user — the catalogue is far too large. Instead the pipeline runs in two stages. Retrieval narrows billions of candidates to a few hundred using cheap, fast similarity: items related to what you engaged with, items popular with users like you (collaborative filtering — “people with your pattern also loved…”), and items matching your explicit follows and interests. Ranking then spends the expensive computation on those few hundred: a deep model predicts, for each candidate, the probability you will click, watch to completion, like, share, comment, or hide — and combines those predictions into a single score weighted by what the platform values (watch time, usually, above all). Freshness, diversity rules and policy filters adjust the final order. What you experience as “the algorithm’s taste” is this scoring function expressing one company’s definition of a good session — which is why every platform’s feed has a distinct flavour even on identical data.

The Feedback Loop: Why Feeds Narrow and Intensify

The system’s deepest property is that it trains on the consequences of its own choices. It shows you what it predicts you will engage with; you engage with what you are shown; that engagement becomes training data proving the prediction right. Around this loop, three effects grow. Narrowing: interests outside the model’s current picture of you are rarely tested, so the picture stops updating — the feed becomes a portrait of your past attention, not your present curiosity. Intensification: because stronger emotional content (outrage, awe, envy) reliably produces engagement signals, ranking drifts toward it — not from malice, but from the objective function doing exactly what it was told. And creator adaptation: producers study what the ranker rewards and manufacture more of it, so the supply itself bends toward the loop. Understanding the loop reframes the common complaints: the feed is not (usually) censoring or promoting viewpoints by hand; it is an engagement mirror with a memory, and its distortions are the predictable physics of the objective it was given.

What Platforms Optimise Is a Choice, Not a Law

It is worth stating plainly that none of this is inevitable, because the objective function is chosen by humans. A ranker can weight long-term satisfaction surveys over raw watch time, can inject diversity deliberately, can demote borderline content, can offer a chronological mode — and platforms adjust these dials constantly under commercial, regulatory and reputational pressure. The same scrutiny applies to darker adjacent systems: the detection mindset in our companion piece on how to spot a phishing page — asking who built this, what does it want from me, what is it optimising — is exactly the literacy these systems require. Regulation increasingly demands transparency about the main parameters of ranking, and the user’s own settings (interest managers, “not interested” buttons, watch-history controls) are real levers, deliberately provided and widely ignored.

Retraining Your Own Feed: a Practical Protocol

You cannot rewrite the model, but you are its training data, and training data has power. Behave deliberately for two weeks and the portrait updates: watch the things you wish you saw more of to completion; use “not interested” and hide functions ruthlessly — they are the strongest negative signals you own; search actively for new interests instead of waiting to be fed them; diversify follows across the boundary of your current taste; and take conscious breaks from the outrage economy, because the model reads your pauses on inflammatory content as appetite. Children deserve a plain-language version of the same conversation: their feeds are trained by their pauses, and knowing that at twelve confers an immunity that platform settings alone never will. Periodically clear or pause watch history on platforms that allow it — the reset button is the closest thing to a fresh identity the system offers. Knowing how recommendation algorithms work does not free you from them, but it changes your position inside them: from product of the loop to one of its authors, feeding the mirror better evidence about who you actually are.