Genre is a discovery hack. If you listen to "Reggae", that may be only because you know you like Bob Marley and the Wailers, but you don't necessarily already know that Peter Tosh plays similar music, or that Desmond Dekker is similar, but is actually considered Jamaican (First Wave) Ska. It won't be able to predict if you will also like or dislike Rage Against the Machine or Elvis Presley, or individual songs that an artist plays out of their normal genre.
If Spotify can solve discovery, by using actual human listening patterns, playlist contents, machine classifiers, or whatever else they may have at their disposal, it is much less important that they have licenses for all the most popular songs. The popular songs are the easiest way for the system to determine what kind of sound any given listener will like, but tuning the radio and auto-playlists to respond to likes, skips, and dislikes should be able to reveal the songs that people will like, from artists nobody has heard of [yet].
I think they're almost there. When I go to radio mode based on my "Liked from Radio" playlist, I want to also be able to specify between "I'm passively listening, so just play music I probably won't want to skip" and "play only new stuff, so I can actively train your algorithm". As it is, it mainly just plays things that are already on my list, and I have to skip ahead on everything I have already liked when I'm trying to train it.
The key issue is that if Spotify is the music discovery engine, people won't discover music that isn't on Spotify. You have to license to them, or you don't acquire new fans. It only helps them that broadcast radio has consolidated itself into a uniform ball of only the greatest hits by only the biggest stars.