Similarity plays a key role in defining the classes of sounds that interact in phonology (Kaun 1995; Wayment 2009). Moreover, a definition of similarity is crucial to the evaluation of both input-output and surface correspondence in Agreement-by-Correspondence (Rose & Walker 2004). However, the representational structure necessary to compute similarity is often left undefined. The current paper encodes similarity as a weighted featural similarity lattice in Gradient Harmonic Grammar (Smolensky et al. 2014; Smolensky & Goldrick 2016). The gradient representational system proposed correctly predicts the behavior of stops in Lezgian, providing a reanalysis of the consonant agreement patterns in Ozburn & Kochetov (2018).