if there no related tracks found, then similar artists are checked for, then a random similar artist is picked and a random top track from the similar artist is picked.
I think you've read what I've said too fast. The problem is when the last.fm node script can't find any similar artist, with the help of last.fm statistics.
However , there's still other artist for the same genre in library.
Example: you play a track from "psyclon nine". Let's say that in the looooong list of similar artists (hey, 25 pages at last.fm) , the script doesn't manage to find a single similar artist in your library (well, because I just don't collect much of that kind of music). So, then if the script look to the top tag associated to artist i.e "industrial", there's much more trivial choices that appears on the top artists like NIN , or Marylin Manson (and much bigger probability to find them on library).
I hope it's clear.