Wired ran a piece on how an early rating site shaped the internet we still use. It is a good reminder that the feed deciding what you see today descends from some fairly simple ideas, and those ideas are worth understanding because they explain most of what feels strange about modern platforms.
The core trick: you are a set of coordinates
The dominant technique is collaborative filtering, and the insight behind it is that you do not need to understand content at all to recommend it.
Build a giant table of who engaged with what. Then factorise it, which in practice means finding a compact set of numbers for each user and each item such that multiplying them reproduces the observed behaviour.
Those numbers are latent factors. Nobody labels them, and they frequently do not correspond to anything a human would name, but they capture real structure. If your numbers sit close to another user’s, whatever they engaged with and you have not seen becomes a recommendation.
The system does not know what a film is about. It knows that people positioned near you watched it.
What it is actually optimising
This is the part that explains the most.
A recommender predicts engagement, because engagement is what can be measured. Watch time, clicks, replays, comments, shares. It does not predict whether something is true, good for you, or something you will be glad you saw.
Those diverge in a specific direction. Content provoking anger or indignation reliably generates engagement, so it is promoted, not because anyone chose to promote outrage but because the metric cannot distinguish it from interest.
Every unpleasant property of feeds follows from this one substitution: the measurable proxy replaced the thing actually wanted.
Why it feels like it is reading your mind
The common suspicion is that phones listen to conversations. The prosaic explanation is usually stronger.
These systems know your location history, who you interact with, what you linger on without clicking, what people near you engage with, and what people similar to you did last week. If a friend you see regularly researched something, you and they occupy nearby coordinates, and the recommendation follows.
There is also survivorship bias in the impression. You notice the uncanny hit and forget the hundreds of irrelevant items you scrolled past that day.
Exploration, and why your feed sometimes goes odd
A recommender that only shows what it is confident about learns nothing new and eventually traps you in a narrowing loop.
So systems deliberately inject uncertain items to gather information. That is why a feed occasionally serves something baffling: it is testing a hypothesis about you.
It also explains how a single click can reshape a feed for weeks. A strong signal in a sparse area moves your coordinates a long way, and the system then explores that region.
Practical control
- Use “not interested” deliberately. It is a much stronger signal than scrolling past, which reads as neutral or even as attention.
- Watch what you linger on. Dwell time is measured whether or not you tap, so pausing on something you disapprove of trains the system towards it.
- Do not engage to argue. A furious comment is engagement, and the system reads it as success.
- Clear watch history where offered. It resets your coordinates rather than nudging them.
- Use chronological or following-only views where they still exist, which bypass ranking altogether.
- Subscribe directly to sources you value, by newsletter or RSS, so the material you actually want does not depend on ranking at all.
The mental model worth keeping is that a feed is not a window onto what exists. It is a prediction of what will keep you there, and those are different objects that happen to look identical.
Join the discussion