It's not even a new algorithm. This technique is known as
"market basket analysis": which things appear together in a
shopping basket, where a basket may be accumulated over several
shopping stops? (Another reason why Amazon has your identity.)
I have no information whether this is the case, but the most
plausible assumption seems to be the following.
If you interpret buying a book as citing it, then the resulting
weighted network of books bought together is the co-citation network
known from bibliometrics, and the lists of frequent co-buys are simply
the neighbors with the heaviest ties to the book at hand. This would
also explain the asymmetry (a tie may be the heaviest for the node at
one end, while its alter may have even heavier ties).
Obviously, there is no way to check for biased weighting schemes
that, e.g., help promote high-margin books.