I was being a bit facetious on the "sophisticated heuristics" part. And I don't live alone. I have plants. You've met them.

Here's more of what I mean. When they're doing comparisons, they're looking for similarities (with the same error-ridden N-dimensional similarity methods I was citing earlier). Upon finding similarities, they look at trending statistics (Human1.propertyX AND Human2.propertyZ imply SUCCESS with N% confidence level). They find these trends based on their previously determined relationship success data, but with anything statistical, it's only to some degree of confidence.

But here's where it bugs me a bit. The things they trend for isn't necessarily "SUCCESS." Not only can they rarely define such a thing (What constitutes a successful relationship? Weeks? Months? Marriage? etc.), but it's super hard to say whether people will be giving them accurate data after the fact. (www.pbfcomics.com)

Instead, the sites I've (briefly) looked at were trending/matching on "compatibility." Which sounds great in an advertisement, but what basis "compatibility" is determined from is never really revealed.

Maybe their process is reviewed by professional psychologists. Maybe by professional relationship counselors. Or somebody who would have any relevant knowledge on the subject. But if you're competing for the repeated business of lonely people, it's not necessarily in your best interest to generate successful relationships. And I bet a doctor's consulting fees are far greater than those of a programmer, or statistician.

EDIT:

I have also yet to see proof that geographic proximity isn't a better contributing factor than anything the dating cites trend on.