In this work, we evaluate the current limitations of iterative similarity searches and investigate several strategies to improve their performance. To evaluate the reliability of iterative approaches as homology inference tools we create RefProtDom—a benchmarking dataset of diverse query domains and full-length proteins containing their homologs in a variety of architectures to simulate searches of real proteins with complex homology relationships. RefProtDom’s homology annotations and boundaries were manually supplemented using local and semi-global searches, reciprocal searches, and structural classifications to ensure accurate alignment evaluation. Using RefProtDom, we identified a previously unrecognized source of error in PSI-BLAST (Position Specific Iterated BLAST) that is responsible for its profile corruption: Homologous Over-extension (HOE). HOE accounts for the largest fraction of the initial false positive errors (hard queries: 86%; sampled queries: 68%), and the largest fraction of false positives at iteration 5 (hard: 51-91%, sampled: 49-69%). We implement a (noExt) strategy to reduce the HOE error that increases PSI-BLAST’s specificity 4-8 fold. We also show that HOE is not a similaritymeasurement or statistical error, but rather an alignment strategy error to which all iterative similarity-searching methods are susceptible. The superior scores and gap penalties afforded by the rigorous strategies of PSI-SEARCH noExt and JACKHMMER, respectively, did not provide any improvement over PSI-BLAST noExt. Rigorous strategies indeed provide better sensitivity in pair-wise (e.g. SSEARCH: 6.2% hard family coverage; BLAST: 4.3%) and better specificity in profile (e.g. PSI-SEARCH: 0.05% hard non-homologous errors; PSI-BLAST: 2%), searches for the unmodified methods. In the absence of a noExt modification, JACKHMMER outperforms PSI-SEARCH, and both outperform PSI-BLAST. But, HOE is such a pervasive phenomenon for all iterative methods, that addressing it is enough to level the methods’ performance. Our noExt strategy effectively prevents the propagation of the HOE alignment pathology, but it does not prevent it from happening when a homology is first identified. We conclude by outlining future directions to improve alignment accuracy, which, if addressed, should allow the sensitivity of iterative searches to more closely approach that of structural comparisons.