The best service, and it won by some margin, was Diffbot (www.diffbot.com). I ran comparisons between approx 20 different services and libraries and it won by some margin. It uses machine learning rather than regular expressions or per-site filters, and the engine has been extensively trained (I threw a lot of edge cases at it, which improved it). There seem to be a lot of similar services that do well with common cases but completely fall apart when applied broadly.
So to the author of this service - what features or examples do you have that distinguish your implementation from others? What is the technique being used here?