The Week You Watched One Too Many Documentaries
It's a Friday night and your homepage looks like a stranger's. The horror films you'd been meaning to revisit are gone from the top row. The quirky comedies you rated highly two years ago have sunk somewhere past the second scroll. You didn't change your taste. You just watched three true-crime documentaries in a row last week, and the platform made a decision about you.
Not a conscious one. No engineer typed your name. But the model updated, the weights shifted, and the algorithm committed.
This is one of the most underappreciated things recommendation systems do: they don't just learn your taste. They freeze it.
The Engine Under the Hood
Most major streaming platforms run on some variant of collaborative filtering, often blended with content-based signals. The basic idea is elegant: find users whose past behaviour looks like yours, see what they consumed next, surface that. If ten thousand people who loved dark Scandinavian thrillers also binged a particular South Korean series, you'll see that series promoted. You never asked for it. You don't need to. Your doppelgangers already voted.
The wrinkle is recency weighting. Because platforms optimise for engagement right now, recent behaviour counts for more than old behaviour, and not by a small margin. A ratio of roughly 70/30 in favour of the last thirty days over everything prior is a conservative estimate for how aggressively some systems discount history. Spend a fortnight in a documentary spiral and the model places a heavy thumb on that side of the scale.
This is where genre burial starts. Not with a ban. With a probability.
Content tied to your older patterns gets assigned a lower predicted click-through score, which means lower placement on the grid, which means you're less likely to see it, which means you don't click it, and the system reads that absence as confirmation that you never wanted it. The genre doesn't disappear from the catalogue. It just quietly sinks past the fold, past the second scroll, past the point where most people ever look.
A Tale of Two Accounts
Take two people, Sofia and Marcus, who both sign up for the same music platform on the same day with nearly identical listening histories: a deep love of post-punk and a secondary interest in jazz.
Sofia keeps a consistent diet. She listens to both genres most weeks, and the algorithm models her as genuinely bimodal, surfaces new releases in both lanes, keeps her discovery queue eclectic. Three years in, she's found artists in both worlds she'd never have encountered otherwise.
Marcus goes through a long work crunch. For about six weeks he leans on ambient and lo-fi playlists because they help him concentrate. The system registers this as a preference shift. His jazz recommendations don't vanish overnight, but they start appearing in the fourth or fifth row instead of the first. He doesn't scroll that far, so he doesn't play them, so the system confirms its model. Within four months, his Discover Weekly playlist has essentially erased the post-punk decade of his life. He still loves it. The platform no longer believes him.
The difference between Sofia and Marcus wasn't taste. It was one busy month.
The Part Most Guides Skip
Explicit feedback and implicit feedback are not weighted equally, and implicit almost always wins. That's the mechanism most people miss, and it matters enormously.
When you click a thumbs-up or a five-star rating, that's explicit. The system notes it. But when you play a song three times in a row, skip the next seven, rewind a scene, or let a show autoplay at 1 a.m. while you're half asleep, that's implicit. Implicit signals are continuous and voluminous, so they carry more statistical weight in most production systems. You cannot out-rate your own behaviour. You can heart every jazz track on Spotify, but if your play counts say lo-fi for six weeks straight, the play counts win.
So here's the question worth sitting with: how many genres have you "liked" your way through while your actual listening told a completely different story?
People think the fix is to rate things more carefully. It isn't. The fix is to actually play the thing you want recommended back to you. Engagement is the currency. Ratings are a tip jar.
What People Get Wrong (And the Caveat Worth Keeping)
The popular narrative treats this as the algorithm being malicious, or stupid, or both. It's neither, and that framing makes people worse at dealing with it. The system is doing exactly what it was designed to do: maximise the probability that you engage with the next thing it shows you. If your recent behaviour predicts documentary engagement at 80% and horror at 30%, surfacing documentaries is the correct call by the only metric the system is accountable to.
Still, there's a real caveat worth keeping. These systems are genuinely good at discovery within a genre once they've locked onto one. If you really have moved on from something, the algorithm is often right. The problem isn't that it updates. The problem is that it updates irreversibly on short windows of data, with no mechanism to ask you whether the shift was intentional. It treats a stressful month like a personality transplant.
Some platforms have started building in explicit taste-reset tools (Spotify's genre mood controls, Netflix's "not interested" button) precisely because engineers noticed users felt trapped. Those tools help. They work against the grain of a model that really, really wants to believe your last thirty days are your true self, which is a bit like judging someone's entire personality by what they ate during a bout of flu.
Getting Your Genre Back
If you want to pull a buried genre back to the surface, the approach is blunt and it works: play it deliberately and completely. Don't skip. Let tracks or episodes run to the end. Do it across several sessions in a short window, because recency weighting cuts both ways. You can exploit it as easily as you can fall victim to it.
Three albums back to back, no skips. Give it a week.
You're not gaming the system so much as feeding it accurate data it stopped collecting.
The deeper point: recommendation algorithms don't have a theory of you. They have a model, and models are only as good as their inputs. A model built on six weeks of lo-fi focus music knows nothing about the decade of post-punk that preceded it. It knows what you played last.
Your taste is not a phase. The algorithm just needs reminding.