Hi @sbombin,
The reason why "normalization" isn't recommend for ANCOM is that it is a compositionally-aware tool which uses relative abundances as input, this in a way is a type of normalization on itself. For other tests like ADONIS/ANOSIM you certainly want to be doing some sort of normalization (like rarefying, CSS. etc). That being said, be mindful that there are compositionally-aware distances also, such as the output of q2-DEICODE. So, if you are using something like Bray-Curtis, UniFrac, etc. you want to perform some normalization, but if normalization has already been address, such as in DEICODE, then you don't need to.
This paper, though a bit outdated by today's standards, might provide some insight into the problem.