21 bytes extra at 1T requests a day = 21TB/day. Probably nothing?

You're correct, at global scale it's still peanuts (and of course it's not a single server but redundant clusters of servers).

I have a problem at your math though, I'm guessing with the overhead and the pretend-IE headers it could be a whole Ethernet packet (which can be up to 1,500TB/day, but realistically it could be around 500TB).

My maths fine, I was comparing it to the comment suggestion of sending "1" instead of "Microsoft Connect Test" as the response body. The headers should be the same.

But I guess you could say the content length header would be 1 byte longer in the double digits length contents (which it is).

In the bigger picture:

  curl -is http://www.msftconnecttest.com/connecttest.txt | wc -c
  520

499TB/d vs 520TB/d, would be that a significant difference?

HTTP is also pretty inefficient. I guess if you went for some sort of custom protocol it could be blocked by firewalls or whatever.