Normalizing put values on a (somewhat) universally coherent value range. For example, --normalizeUsing RPGC will normalize the sample(s) to 1x coverage before proceeding. Scaling is done when you have 2 samples and needs to be done to make them comparable (i.e., it corrects for differences in sequencing depth).

• Using normalization whenever you need to compare multiple samples, or when it's the relative change in signal that matters (e.g., ChIP-seq)
• Scaling alone should be used when you need to compare samples but need the output to mimic the raw read densities. This is often useful for things like ribosomal-profiling.

For ChIP-seq stick to normalization.