There are billions and billions of pages in the index, let's say they produce a page classifier that can correctly identify spam an unrealistic 99% of the time.

There are billions and billions of pages in the index, let's say they produce a page classifier that can correctly identify spam an unrealistic 99% of the time. They'd still misidentify tens of millions of pages. Regarding your second point, how do you determine usefulness in a way that can't be gamed? My point is, this is a very, very, very hard problem to solve at their scale, a lot of posters on here are trivialising it.