Test one pair at a time on a random 20 percent sample, wait 4 hours, then send the winner to the remaining 80 percent. Record what you learn for future sends.