Analysis of CASP2 protein threading results shows that the success rate of structure predictions varies widely among prediction targets. We set "critical" thresholds in fold recognition specificity and threading model accuracy at the points where "incorrect" CASP2 predictions just outnumber "correct" predictions. Using these thresholds we find that correct predictions were made for all of those targets and for only those targets where more than 50% of target residues may be superimposed on previously known structures. Three-fourths of these correct predictions were furthermore made for targets with greater than 12% residue identity in structural alignment, where characteristic sequence motifs are also present. Based on these observations we suggest that the sustained performance of threading methods is best characterized by counting the numbers of correct predictions for targets of increasing "difficulty." We suggest that target difficulty may be assigned, once the true structure of the target is known, according to the fraction of residues superimposable onto previously known structures and the fraction of identical residues in those structural alignments.