[almostnormal]

If an author is unable to produce readable commits on the branch of a PR the author is likely not able to produce readable change-sets either.

To review I like branches. Bad commits are not pleasant to look at, but they carry information, too, e.g., revealing the need for discussion. It's a bit like asynchronous pairing.

I don't think looking at the code on a website is sufficient as review in many cases, so I typically have a copy of the branch anyway. If any feature is needed, it is the possibiliy to comment on unchanged lines.

It probably depends on the scope. If the reviewer works on a more abstract level and can rely on authors to get the details right, maybe github is not ideal. Where the reviewer is almost as deeply involved as the author, a branch works nicely.