Nonce word studies motivate a notion of gradient similarity between nonce words and real words. In morpho-phonological research, similarity is often taken as to be a relationship between a nonce word and the list of morphemes / words that undergo a given morphophonological alternation (Albright and Hayes 2003; Becker et al. 2011 i.a.). This paper challenges this view on the basis of nonce word data on Russian vowel–zero alternations (Gouskova and Becker 2013; Becker and Gouskova 2016). I propose a model where morpho-phonological similarity is a relationship between the available underlying representations and the underlying representation the nonce item must have in order to undergo the alternation. The implementation of the proposed model matches—and in some comparisons exceeds—the performance of Becker and Gouskova’s (2016) MaxEnt-model. This study thus presents a linking hypothesis between nonce word studies and approaches that mark segments themselves as undergoing certain restricted alternations.