Alpha-to-coverage models so-called screen door transparency. Basically, when you look at an object like a wire mesh from a distance, it will look transparent even though it consists of opaque material with holes. Alpha-to-coverage uses multisampling to implement just that on a sample level: If you use 4x MSAA, you can render a 50% transparent object by masking out two samples per pixel, and writing the other two samples without blending.
Because samples are either fully opaque or masked, its characteristics for depth sorting are similar to alpha test using a dithered alpha value.
