Based on their definition it should be counting the entire IP packet, which would include the header.
But since the IP packet may be layered inside of a transport protocol, or delivered as the payload inside of an ethernet frame, it sounds like there may technically be some headers or trailers that wouldn't be counted.
All the technicality aside, it is probably a pretty good measure of actual bandwidth utilization but there are some things listed in the manual that elude to reasons the numbers may appear to be inaccurate.
Some of them are listed in man pages at the link above if you scroll down to the "Quirks(aka they're features not bugs)" section.
I think the main thing to keep in mind is that the bandwidth is being calculated using packets, and a packet includes the IP header. A frame may or may not include an entire IP packet, and whether or not the frame header and footer add a significant amount of overhead to throughput or bandwidth utilization is somewhat non-trivial and I am personally not sure.