Artificial grammar learning (AGL) has been used extensively to study theories of learning, but compelling conclusions cannot be drawn without an analysis of individual strategies. We describe a new statistical method for doing so, based on the increasingly popular framework of latent variable models, which is especially suited to capture heterogeneity in participants' responses. We applied the method of latent class regression models, in which the intercept and regression coefficients can have different values in different latent groups of participants; each latent group represents different reliance on the potentially available sources of knowledge in AGL, such as grammaticality and fragment overlap. The results indicate that grammaticality and fragment overlap can be understood as distinct aspects of learning performance, as evidenced by different groups of participants adopting predominantly one or the other strategy in a series of comparable datasets from AGL studies.