Actually the bounding box for snapping is computed from all the current selection but in my case it will be smarter if it only use the bounding box of the main item of a selection when multiple items are selected. And to return to the way it works actually, I will just have to group the current selection to have a single object.
See attachment for an example:
I just want to snap to the middle bounding box of the "middle man" not the bounding box of all the selection. In the case I want to snap to all the selection, I will group it.