I never knew A can be like B without B being like A.

The matching score is probably the same, or very close in both ways, but this fact does not necessarily help in a three-way scenario:

    A <-> B: 80%
    A <-> C: 90%
    B <-> C: 70%

When you search for A the best match will be C, but if you start with B it will be A. If one of the accounts has a smaller sample set as in GP's case, the gap could be quite big.

I'm still in disbelief. I think one should run the operation in reverse after obtaining result set.