Shaddows - 4:18 pm on Apr 8, 2011 (gmt 0)
I'm saying that multivariate testing happens. I'm saying that profiled traffic sets are given profiled SERPs (organised site sets), and the resulting user behaviour is fed back into the system. There was a thread about it a while back.
You would do this to test SERP satisfaction (NOT site or page specific) as well as split testing the specific results. Bounce rate would be one of many variables- but only to be interpretted in context of the traffic fed, against expected norms for the page/site profile.