imho, that page is a joke.
example:
qwertasdfgyxcvbn = 18%
now replace every third character with a "1":
qw1rt1sd1gy1cv1n = 92%
yeah, right...
That comes from the classification of characters, of course that's not exactly accurate. But there is no better way, the randomness of a string depends on the set of possible strings you assert it comes from.
If you assert that it is letters only, you have way less passwords to try, then if you assert that it is letters plus numbers.
The site seems to check only for that, which I mentioned in the OP depends on the assertion that all characters are independently chosen with equally distributed probability.