From a purely statistical standpoint - their claim is false about the roughly 1 in 16,777,216 chance. But the goal is to find like-minded people.
Let's create a 1 question True/False test. Let's assign the probability of answering True on the question is 70% and answering False is 30%.
You give the test to 100 people. You now have 35 pairings for "True" and 15 pairings for "False". Would it make sense to pair the "True" people with the "False" people to approach an "even distribution"? Only if your goal was to match people with "1 out of every 2 people" from a purely statistical standpoint of weighing 2 options.
The importance in this case is not distribution - but rather if these 8 questions determine who is "similar" in thinking to another person. If you asked someone their favorite color and their favorite pet - you might get a lot of matches. But many of those matches might be terrible with the people having little in common beyond that.
In the FAQ they take a step back and don't really guarantee good matching, even in the event of a match. However matches should be statistically rare (even if biased towards a specific 8 answers on the questions) and might still produce "good results". This makes it an interesting case study and one that can be tweaked and redone if we ever discover a way of asking only a few questions (I'd say no more than 10?) and accurately defining someones personality.
That's the way I see it at least. More of a case study than a statistical claim, even if they're trying to spin the statistics to spur people to take the 8 question quiz. Who knows, it may end up with a lot of good matches based on just-vague-enough questions and the likilihood of similar answers (even if, in some scenarios, fewer than the projected 16~ million and in other scenarios even more)