Two basic approaches, combined
Collaborative filtering recommends based on what similar viewers watched — if people with a viewing history like yours also loved a particular show, it gets recommended to you, without the system needing to understand anything about the show's actual content. Content-based filtering works the other way: it recommends based on attributes of what you've already watched — genre, cast, pacing, themes — independent of what anyone else did. Modern streaming platforms blend both, plus dozens of secondary signals (time of day, device, how much of a title you actually finished versus abandoned).
Completion rate matters more than you'd think
Clicking play tells a recommendation system much less than finishing an episode does. Abandoning a show ten minutes in is a strong negative signal, even stronger than never clicking it at all — which is part of why a slow-burn show with a great later payoff can struggle to get recommended to new viewers even when the people who stick with it love it.
The thumbnail is part of the algorithm
Streaming platforms routinely test multiple thumbnail images for the same title and show different viewers whichever image their data suggests is more likely to get that specific viewer to click — meaning two people can see completely different cover art for the identical show, personalized not just by taste in genre but by what visual style tends to catch that viewer's eye.
Why the algorithm sometimes feels stuck in a loop
Recommendation systems can create feedback loops: if you're recommended mostly one genre, you mostly watch that genre, which reinforces the system's model of your taste, narrowing future recommendations further — sometimes called a "filter bubble." Deliberately watching outside your usual pattern, or using a platform's genre browse pages instead of the algorithmic home row, is one of the few reliable ways to break out of it.
What it can't tell you
No recommendation algorithm currently understands why a show is good the way a human critic does — it optimizes for continued watching, not for artistic merit, which is why algorithmically-boosted shows and critically acclaimed shows are often different lists entirely. Treat the algorithm as a discovery tool, not a quality signal.