I don't think the gTrends argument works though (see http://news.ycombinator.com/item?id=4741590). Most of the results inserted are super-recent (and so I expect the recency of the trends data -- like the last 90 days -- to dominate). Of course it is all a black box so who knows.
Some of the more interesting things to me are:
--you can't reliably de-personalize (as you cited).
--the variation in results across our study was great.
--variation for signed out (even incognito) users was not much different from signed in users.