Of course it is a matter of metrics - it has no way of knowing what is useful. The closest way to algorithmically discover (outcomes over time) would be prone to spurious correlations and be so intrusive it would make Cambridge Analytica look like LavaBit.
>>Nasrud+(OP)
I'm thinking "make things more discoverable" than "find more useful things" if that makes sense. I'm willing to wade through it myself if you present me with options.