It's easy to poke a hole in a methodology, but this was a pragmatic short curcuit of collecting websites as they are in the wild. You could ask 10 experts to code up equivelent websites and measure performance and still someone would complain that it's a synthetic benchmark. The nice thing about this methodology is that noone was expecting the dataset to be used for this purpose. It is an interesting set of measurements in my opinion.

Edit: the biggest issue is that the article is vunerable from severe confirmation bias from the authors admitted sunken costs