We compare element trees directly. It would be easier to extract the text and just do a regular text diff, but this loss of element/tree info means you can't reconstruct html versions of the page with differences highlighted, which for us is an important feature.