It's mostly in the context of tile-based renderers vs immediate mode renderers ...
If we assume a naive traditional forward rendering pipeline, all of your fragment shaders are certainly paying for overdraw cost on IMR/PC GPUs but there's much more twists and turns to be had. With deferred rendering, we can avoid paying some overdraw cost as we use our G-buffer to reject shading occluded fragments during the lighting pass. Overdraw will occur during G-buffer generation. If we apply a depth-only prepass or a visibility buffer for either cases then overdraw does not become a concern at all during fragment shading but you now have to render 2 geometry passes which will raise costs elsewhere in the pipeline ...
One of the advantages behind tile-based architectures is that it can perform complete hidden surface removal without needing an extra geometry pass hence why they have virtually no overdraw cost during fragment shading. On all mobile GPU IHVs, this is the reason they tell the developers not to ever do a depth prepass ...