Is it? I do not know of any paper who has shown this and I do not recall to come across one that in turn showed that GC normalization was beneficial. That was recently asked on SE was well, https://bioinformatics.stackexchange.com/questions/15006/gc-normalization-atac-seq-data and Devon Ryan there answers pretty much the same I would say as well. GC bias normalization has been explored and demonstrated to be beneficial in some cases towards RNA-seq, see this paper from Mike Love (that is the DESeq2 author/maintainer), but not for ATAC-seq. I personally recommend, even if you find a paper who does this, not to adapt these strategies unless there was a credible biostats paper to really investigate this and its implication on the data. It is not commonly done for ATAC-seq (nor ChIP-seq or any other DNA-seq) and I would actually recommend to skip it rather than blindly applying any non-standard normalization. You would probably need to normalize for the GC content of every read/fragment I guess since a peak does not necessarily have the same GC as every read...not trivial all in all. I gave my two cents on how to (relatively simply but properly) normalize ATAC-seq data for something like bedGraph/bigwig-like files in A: ATAC-seq sample normalization (quantil normalization) and for every differential analysis one commonly uses established tools like edgeR, DESeq2 or limma.