As the above, it's about recommending unplayed games to you that you already own.
My point of criticism is that it seems to be based on similarity - ie finding games that score the same across a bunch of factors (with the exact correlation determined by ML). However my game choices are as much determined by playing something that's completely different to what I've just played because I want a change. If the ML is trained based on my own play patterns only then it might work (though I doubt there's enough data). If it's trained on the community at large then it'll fail.