Screen reader
Does not read the screen but the accessibility tree — which is why a beautiful page can still be unusable.
A screen reader converts the interface into speech or braille. JAWS and NVDA on Windows, VoiceOver on Apple's platforms, TalkBack on Android. What it reads is not pixels but the accessibility tree, a structure the browser builds from the semantics of HTML elements and any ARIA attributes.
Users rarely navigate from top to bottom. They jump between headings, landmarks, links and form fields, which makes heading order more important than visual order. The common failures follow from that: a <div> with a click handler that is never announced as a button, images without alternative text, and ARIA attributes written on top of correct semantics that make it worse. Properly built HTML needs almost no ARIA at all.