why even bother ?
So why did they do this ?
The only reason I can think of is that they want to use stereotypes in one form (email addresses etc) to estimate how to use stereotypes in another (the types of advert that is intended to appeal to stereotypical men and women). Seems like it's truncating precision as far as it can by reducing to binary gender, then extrapolating from that to choose an advertising strand.
It difficult to think of a less useful thing to do.
A more useful thing that doesn't pre-assume the results might be to attempt to link the input data with the desired outcome : If firstname.lastname@example.org actually BUYS dresses, that's a useful result that doesn't just throw data away,