w3counter is totally unreliable. The source of their data is just the hits to their own website.
netmarketshare and statcounter results are from their web analytic products which span across thousands of websites. While their results can still be debated over various aspects, I think that netmarketshare is more reliable of the two.
statcounter approach is fundamentally flawed in that they rely on individual use page loads. Each page load is seen is a unique user hit. For instance if an IE user visits one page of a website, Firefox user hits 3 pages of the same website and chrome user hits 6 pages, statcounter will show the browser market share as IE= 10%, Firefox = 30% and Chrome = 60% when realty, there was only 1 user for each browser.
netmarketshare post-process their data and count unique users across pages of a website. So even if the same chrome user hit 6 pages of the same website, he would be counted as a single user.
CORRECTION: w3counter seems to be operating in the same method as statcounter except that since they rely only on a subset of total page hits,its even more unreliable.