My gut tells me it's akin to doing principle component analysis. To find the most differentiating question, do 1 dimensional PCA, and the question most aligned with the resulting dimension is question #1. Then do 2 dimensional PCA fixing the first dimension as the first question. The result is question #2. And so on.
OF FAR GREATER CONCERN: They are collecting email addresses associated with personality questions. There is no privacy policy, other than the statement "If there's a match, we'll email you. (You will not be emailed under any other circumstances: no promos, no newsletters, nothing except news of a match.)" You don't know to whom you are giving this information, as there is no named legal entity, just a gmail address and a domain name whose registered owner is hidden in whois databases.
But, even given all that data...I find about half of my 99% matches (of which I've talked to maybe a half dozen and gone on a date with maybe four) aren't particularly compatible or interesting, to me. There seem to be other factors at play that just aren't measurable. And, heck, maybe being just like someone else isn't what makes an ideal friend/lover/partner.
Or a decision tree, maybe
I wonder if you couldn't do better with binary questions. I'm reminded of the site http://www.correlated.org/.
Suppose you were to ask a bunch of binary questions, and then use some statistics to find questions that have fairly evenly distributed and non-correlated results, then take 24 of the best questions by those criteria (arbitrary, but 2^24 = 8^8) and find matches.
So basically, ask a whole bunch of binary questions, find ones with roughly 50-50 answers that don't correlate with each other, make the match-ups based on those and ignore the others.