Seed propagation-based stereo matching can help to reduce ambiguity occuring when a pixel from one image has different putative correspondents in the other one due to difficult areas (repetitive patterns, homogeneous areas, occlusions and depth discontinuities). They rely on previously computed matches (seeds) to reduce the size of the search area, and thus the number of candidates. One approach of these iterative methods selects the “best” seed at each iteration to prevent the propagation of errors. However, little attention has been brought to this best-first selection criterion for which a correlation score is usually employed. This value itself does not consider any ambiguity and is not well-suited to select the most reliable seed. Therefore, in this paper we introduce a reliability measure. It has the advantage of taking into account information from the other candidates, and leads, according to the provided experimental evaluation, to better results than the correlation score alone.