A face model does not see a face. It sees a mesh — commonly 468 landmark points draped across the surface of your head, each with an x, y and z coordinate. Everything else is arithmetic performed on those points.
The first thing computed is a set of ratios: interpupillary distance against face width, nose length against mid-face height, lip width against mouth-to-chin. Ratios matter because they survive changes in camera distance, lens and lighting in a way raw pixel measurements never do.
The second layer is shape classification. Jaw contour points are fitted against reference outlines to sort the face toward oval, round, square, heart or diamond. Most real faces land between two categories, which is why a good report gives you a blend rather than a verdict.
The third layer is expression-independent structure. Models subtract the effect of your current expression so that a polite half-smile does not get recorded as permanently full cheeks. This is the step cheap face-reading apps skip, and it is why their results change every time you blink.
None of this is destiny. It is measurement. What face reading traditions add on top is interpretation — centuries of it, occasionally wise, occasionally nonsense. Knowing which layer you are reading is the whole skill.