Exactly, whether or not it’s the “actual thought” of the model, it does influence its final output, so it matters to the user.
We don't really know that. So far CoT is only used to sell LLMs to the user. (Both figuratively as a neat trick and literally as a way to increase token count.)