I am not entirely sure about when I should use which and what benefits there are.
Because even if you make an entire group of objects in an SOP it can still spit them out as instances.
It seems easier to control the SOP but it also “feels wrong” to make everything as “one object”.
Can anyone explain this a bit? When should I try to make everything in an obj subnet and when should I do it in a SOP subnet (turning the respective ones into HDA's)

Using an obj subnet results in multiple meshes to be created as separate gameobjects, which is expected behavior. If you use a SOP subnet, then only the display node geometry is created, unless you use instancing.

To keep things clear, I'd recommend that if they are indeed separate gameobjects with different behaviour or purposes, then place them in obj subnet. Only use a SOP subnet when they are parts of the same object. I can give a better answer perhaps if you can given an example or describe what you're trying to do.

Thank you.
No it's very adequate. I found it easier to input geometry into a SOP but I can see now, for example, that instancing prefabs with the “unity_instance” attribute requires the use of an OBJ network.
I am working on populating terrain procedurally and did it with a SOP initially.
Then I converted it to an OBJ and have just kept working from there and it has worked very well!