> I count 2^7 = 64 possible variations (counting the madlib/form as just one point of difference), which would take an awful lot of traffic to discriminate between.

A fractional factorial design (http://en.wikipedia.org/wiki/Fractional_factorial_design) of the experiments and other statistic techniques (http://en.wikipedia.org/wiki/Bayesian_experimental_design) may decrease the amount of traffic needed.