Well, I imagine they use metrics like how many pages you sift through before selecting a link (presumably the page you were looking for), or whether you had to modify the search. As some other posters have said, observing a lift in 1% for these kinds of metrics can hide a massive loss of quality for niche queries or user groups. I imagine folks internally don't just look at aggregate stats before pushing changes, but refine metrics based on the type of query or the type of user.