First, are these A) technical or B) biological replicates? That is, the same biological sample run several times with the same antibody (same lot also if polyclonal) protocol, or different biological samples run the same way with the same protocol?
If it is A it may be reasonable to merge them for some analyses, such as just annotating peaks. I would merge the bam alignment files and then do the calls versus merging the calls.
However, first you have analyze your replicates to check they they all perform the same. We did a lot of performance comparisons here:
You can steal those ideas, especially using the ENCODE segmentation tracks if it's human and they have tracks for something like your cell type. But just counting the reads in bins and then doing a correlation is pretty informative.
But even in our data, and we used a robot and do it a lot, one of our technical replicates behaved strangely. See supplemental figure S6.
If it is B, biological replicates, you almost certainly don't want to merge them. You will lose your information about biological variance is present. If you are looking at something like differential peaks between conditions DESeq and really all reputable programs will want some sort of replicates, almost always biological. In general, if you want to compute a p value on anything you need separate replicates (not merged).
If you are just annotating peaks you don't need a p value.