To what extent can we predict how evolution occurs? Do genetic architectures and developmental processes canalize the evolution of similar outcomes in a predictable manner? Or do historical contingencies impose alternative pathways to answer the same challenge? Examples of Müllerian mimicry between distantly related butterfly species provide natural replicates of evolution, allowing us to test whether identical wing patterns followed parallel or novel trajectories. Here, we explore the role that the signaling ligand WntA plays in generating mimetic wing patterns in Heliconius butterflies, a group with extraordinary mimicry-related wing pattern diversity. The radiation is relatively young, and numerous cases of wing pattern mimicry have evolved within the last 2.5–4.5 Ma. WntA is an important target of natural selection and is one of four major effect loci that underlie much of the pattern variation in the group. We used CRISPR/Cas9 targeted mutagenesis to generate WntA-deficient wings in 12 species and a further 10 intraspecific variants, including three co-mimetic pairs. In all tested butterflies, WntA knockouts affect pattern broadly and cause a shift among every possible scale cell type. Interestingly, the co-mimics lacking WntA were very different, suggesting that the gene networks that pattern a wing have diverged considerably among different lineages. Thus, although natural selection channeled phenotypic convergence, divergent developmental contexts between the two major Heliconius lineages opened different developmental routes to evolve resemblance. Consequently, even under very deterministic evolutionary scenarios, our results underscore a surprising unpredictability in the developmental paths underlying convergence in a recent radiation.