It's not based just on text code, but on interactive manipulation as well, so it's not like e.g. TikZ which can be kept in the document source and version controlled, etc.
An analogy would be the weights of pretrained machine learning models. They're still serializable, though it's true they're less semantically trackable than a well-written TikZ snippet (which, by the way, I haven't seen many of in practice!)
It's not based just on text code, but on interactive manipulation as well, so it's not like e.g. TikZ which can be kept in the document source and version controlled, etc.