I've read this scribe that provides a public coin interactive proof for graph non-isomorphism.
In the proof, the verifier samples both a pairwise-independent hash function and a target $y$. Then it continues to show bounds for the probability that there's a pre-image of $y$ w.r.t. the hash function that's also inside a given subdomain $S$.
In all calculations carried out in page 5 (lemma 3), I couldn't find a single instance that used the fact that $y$ is random. It seems as if randomizing the hash function is enough.
This is also the case for any other source I found that shows a specific public-coin interactive proof for GNI.
Why do we need to randomize $y$?