I'm confused by the wording here, it says that "deferring shading shades only visible fragments", but when would normal shading ever shade NON-visible fragments?

Don't we only shade a fragment in the normal pipeline after we know that the triangle will be visible on the screen?

I think the core idea of deferred shading is exactly what you said: only shading triangles that will be visible on the screen after rasterization. Forward shading, on the contrary, will calculate fragments for all geometries and all lights, which may not pass z-buffer and appear on our screen. This is where the speed advantage of deferred shading comes in.

However, deferred shading has several disadvantages as well, like the inability to handle transparency. The Wikipedia page gives a quite helpful explanation https://en.wikipedia.org/wiki/Deferred_shading