Pretty much this. The least obvious at the time, but looks extremely obvious in retrospective. You and your friend are tightly linked on the social graph. Your

Pretty much this. The least obvious at the time, but looks extremely obvious in retrospective. You and your friend are tightly linked on the social graph. Your friend starts watching a new show and keeps googling and watching stuff about it and talking about it with other people online. While the system won't recommend to you every single thing your friend watches and likes, it knows that for some of those the likelihood of your tastes matching is high. E.g., you might not like horror movies unlike your friend, so the algo wont be recommending them to you, despite your friend watching those. But the algo knows you both like crime dramas, so as soon as your close-social-link-connection-friend starts watching and googling and talking about Breaking Bad, the algo thinks you might like it too, so you start getting suggestions for it. It is kind of silly, but it is indeed scarily way more accurate/effective to predict recommendations that way, as opposed to doing very invasive stuff like actually trying to listen to mic or correlate locations or whatever else (not even mentioning that on modern phone OSs it is somewhere between very difficult and pretty much impossible to do in a legitimate way). Tldr: if you have close social connections whose certain media tastes match yours in some specific niches, don't be surprised that when one of those social links starts watching and enjoying something in those specific niches (and googling about it and talking about it to others online and etc.), you might get recs for that type of content too.