Recursion is argued to be the crucial property distinguishing human and non-human primates language learning faculty (Hauser, Chomsky, & Fitch, 2002). Recently, 2 studies (Bahlmann & Friederici, 2006; de Vries, Monaghan, Knecht, & Zwitserlood, 2008), which investigated the learnability of a recursive artificial grammar of the type of AnBn, used the same material but reported divergent results. We propose that the organization of the linguistic environment crucially determines learnability of the recursive structure, and that this factor might offer some explanation to the incompatible findings. In a grammaticality judgment task using the same materials as in Bahlmann and Friederici (2006) and de Vries et al.’s (2008), we found significantly better performance when the training input was arranged in a starting small fashion, than when it was organized randomly.