We have a custom import written for each individual network/feed, so we try to iron out any "quirks" as early as possible.

And for the matching itself, that again is custom-written. I'll be honest - I don't fully understand the maths of it myself (I leave that to the machine learning experts on the team) - but it basically extracts weighted keywords and phrases from your pages, matches them against an index of our products and then alters the weightings as performance data comes in.