Worth a careful read, as is Danny Sullivan's response at Search Engine Land.
As I read the article, the 0.02% number refers to the number of fradulent clicks discovered through client-requested investigations divided by *all* the clicks on Google.
This doesn't mean that 0.02% of clicks are fraud that was not discarded automatically. For that to be true, *all* clicks would have to go thorugh the manual investigation.
In other words, the percentage is calculated by using different scopes for the numerator and the denumerator. The type of classic math mistake one would not expect from Google.
The only valid percentage is the number of manually-discovered fraudulent clicks divided by the number of clicks in those campaigns that complained and were investigated.
Presumably, most campaigns don't complain and thus are not investigated, meaning that they hide some additional clicks that would have been found through manual investigation.
And also interesting that Ghosemajumder's post doesn't address the most central aspect of click fraud -- that almost all of problem occurs on the contextual networks.
For an article in Catalog Success (forthcoming), I had the opportunity to interview many of the expert brains at the reputable click fraud detection vendors.
Those interesting conversations strengthened two prior beliefs: (1) that Google does a very solid job detecting and eliminating bogus clicks from core search, and (2) bad clicks occur in epidemic proportions on the content networks.