When Netflix says that around 80 per cent of viewing on its service starts from a recommendation rather than a search, it is describing the quiet transfer of a decision. You still press play, but a ranking model chose the shortlist, ordered it, and in Netflix's case even picked which still image from the programme was most likely to make you stop scrolling. Understanding how that model works explains a surprising amount of modern viewing habits, including why your suggestions get narrower the longer you stay.
The core method is collaborative filtering: instead of analysing what a film is about, the system analyses who watched it. If thousands of accounts that behave like yours went on to finish a particular series, the model infers you probably will too. Each viewer and each title is compressed into an embedding, a list of a few hundred numbers positioned so that similar tastes sit close together in mathematical space. Your history drags your point around that space every day; the recommender simply reads off what sits nearby.
At the scale of a big platform this happens in two stages. A candidate generator first reduces a catalogue of thousands of titles, or in YouTube's case hundreds of millions of videos, to a few hundred plausible options in milliseconds. A second, heavier ranking model then scores each candidate on predicted behaviour: probability of a click, expected minutes watched, likelihood of completion, likelihood you return tomorrow. Google's engineers described exactly this architecture for YouTube in a 2016 paper, and versions of it run behind Netflix, Amazon Prime Video, Spotify and TikTok. Crucially, none of those predicted quantities is "how good is this" or "what did the user say they wanted". The thumbs you give and the genres you tick at sign-up are weak signals; what you actually watched at 1am is a strong one, and the model trusts the stronger signal.
The choice of target changes the culture downstream. In 2012 YouTube switched its ranking objective from clicks to watch time, because click-optimised systems had learned to reward misleading thumbnails that people abandoned seconds in. Creators promptly restructured videos around retention, with longer runtimes and delayed payoffs, because the metric had changed. Netflix's personalised artwork works on the same logic: the company runs continuous experiments in which the same title is fronted by different images for different viewers, a romance-forward still for one account and an action shot for another, because the model predicts which frame maximises the chance of a play. The programme itself never changes, only the pitch made for it.
Why your feed narrows
The awkward property of these systems is that they can only learn from what they have already shown you. If the candidate generator never surfaces Korean thrillers, you never click one, and the model reads your silence as disinterest. Researchers call this exposure bias, and it compounds: each round of training data is filtered through the previous round's recommendations. Platforms counter it by deliberately injecting a slice of exploratory or randomised content, but exploration costs engagement in the short term, so the pressure runs the other way. The result many viewers notice, a homepage that feels like an echo of last month, is not a malfunction. It is the optimisation working as specified.
Session-level tricks sharpen the effect. Autoplay countdowns remove the decision point at which people previously stopped watching, and next-episode ranking weights whatever keeps the current session alive. Because the models retrain on the behaviour those features produce, the data ends up certifying the design.
What oversight looks like in Britain
Regulators have started treating recommender systems as infrastructure rather than trade secrets. Under the Online Safety Act 2023, Ofcom requires the largest user-to-user services to assess how their recommendation algorithms contribute to the spread of illegal material and, for children's feeds, to configure recommenders so that harmful content is not amplified to under-18s; its codes of practice took effect in 2025. The EU's Digital Services Act goes further for very large platforms, obliging them to offer at least one feed not based on profiling, which is why UK users of some services now see a "latest" or non-personalised option beside the default.
None of this makes the systems sinister. A catalogue of 15,000 titles genuinely needs a filter, and prediction is a defensible way to build one. The practical adjustment is knowing what the machine is actually doing: it is not curating for quality, and it is not listening to what you say you like. It is betting on what will keep you watching, and every bet it wins becomes the evidence for the next one. Search for something specific occasionally, and you move your own point in the space.