They largely use semantic metadata (like alt tags on images). This is part of why the "google docs is moving to canvas rendering"[1] caused a big stir there (accessibility would need to be implemented from scratch, possibly by using a hidden parallel DOM)
