Headers wouldn't work, because a prefetch often turns into a page view.

Edit: their methodology is public. They use the aforementioned API, and headers for the two other browsers that prerender (Firefox and Safari, who rarely prerender because pages must explicitly configure it). Pages pre-rendered and discarded would add 1.3% to Chrome page views if they weren't discounted. Here is the StatCounter FAQ: http://gs.statcounter.com/faq#prerendering

It didn't seem to make any non-marginal effect on the graph. I wonder if they can accurately account for the pre-rendering.

Of course, pre-rendering might just be that good, that it's always hitting the right pages.