The metrics shifted. Subscriber retention just isn't that tightly correlated with them identifying hidden gems for you, that may be an externality we've just lost.
There used to be a lot of competition in this space, for a few years you could go on free sites designed only to give you targeted recommendations. I think Jinni was one.
They were just training their engines and became ad networks.
One other factor, not really anyone's fault -- the massive increase in high quality scripted content has made search less critical. If you fail to find hidden gems in 2003, you get Pirates of the Caribbean sequels and miss out on Oldboy or Goodbye Lenin. Today if you pull up any random streaming service you'll probably find something that's good enough, even if not great.
It's a little disappointing, maybe there's room for a p2p ML recommendation engine in this space for people who really want to optimize what they're consuming. I'd be into it.
edit: i could be mistaken though
Now the recommendations are just not that great. It's like they're not trying too hard because they are comfy in their market position.
I have never watched any anime at all (really not my thing), but Netflix keeps recommending it to me despite the fact that I haven't taken the bait once.
Also it recommends shows to me that it should know through history I've already watched every episode of. Maybe that's on purpose since people watch things again, but I still don't find it useful.
Also it leaves stuff in "Continue Watching" for like a month after I give it a thumbs down rating and remove it from "My List". That should be a strong enough signal to act on.
And it seems to recommend stuff based on stuff that I only "watched" because it auto-played. That should not be a strong enough signal to act on.
The overall point is they seem to have several obvious opportunities to be smarter with content discovery, but they're not taking them.