edited: to respond to abraham's post directly

There's a difference between training with that data and testing against it, which is precisely what I wanted to highlight here since many expected Google to use such data for training. You can't do that since you can overbias/over-fit given what little daat we, plug-in users, are able to produce versus the amounts they have access to.

2nd edit: it should also be noted that that 84% figure they mentioned related to the extension data is a measure of recall. Assuming Google wants to do a better job with precision in these cases (don't want to get many false positives), that recall is still fairly good in mind since I'm sure there's some noise in that data.