Voice similarity score

A voice similarity score is a 0–100 metric comparing a single piece of writing against a brand’s voice fingerprint, used to flag voice drift before content ships.

The score combines structural signals (sentence rhythm, lexical diversity, formality) with judgment-based signals (signature-phrase use, banned-cliche presence). Scores above 75 usually mean the piece reads as on-brand; scores between 50 and 74 indicate borderline drift worth a rewrite; scores below 50 mean the piece has drifted enough that publishing it would visibly erode voice.

Implementations vary in how heavily they weight each signal. Some are pure LLM-as-judge (the model rates the piece); others combine deterministic heuristics with an LLM call. The best surface a score on every draft and a chart of how scores correlate with engagement on the brand’s actual audience.

Without a score, voice drift is invisible until follower engagement starts dropping. With a score, drift is caught before publish — turning brand voice from "we hope so" into a measured variable.