The coordinate analogy is apt, but what trips people up is that unlike geographic coordinates, the "distance" between embeddings shifts depending on what model created them — so you can't mix and match without retraining everything around a consistent space.