Oh but they are. For the vast majority of content creators, semantics are inseparable from presentation; an article I write really does the semantics of looking how it looks, and can't be transformed into something different.

Just like in programming, to extract a good abstraction such as the semantic content of an article you need to test against three different implementations. But most people only write for and test against a single presentation version, so it should be no surprise that even if they try and separate the semantics and the presentation, they'll get it wrong.