“On-brand” is a measurement, not a claim. To benchmark it, score every generated asset against the brand system before it ships, then publish the pass rate as a number you can audit.
Why “on-brand” needs a score
Most teams treat brand fidelity as a feeling. A reviewer glances at a creative, nods, and approves it. That works for ten assets a week. It collapses at the volume generative output now runs at, where one campaign can produce thousands of variants and no human reads them all.
A score closes the gap. Define the brand as a system the model can be checked against: palette, logo placement, typographic rules, voice, claim boundaries, the things a brand guideline already specifies. Test each output against that system and record pass or fail. The result is a rate. A rate is a benchmark, and a benchmark moves.
Keeping a model on-task while it generates is a separate discipline, covered in how to make an LLM focus. This piece picks up after generation: did the asset land on-brand, and can you prove it with a number.
The method: a per-output rubric
A usable rubric scores along four axes, each binary at the asset level.
Visual: the asset uses approved colour, logo, and layout, with no off-system substitutions. Voice: the copy matches the brand’s register and reading level. Claim: every stated number or benefit traces to something the brand can defend. Composition: the asset holds together as the brand would ship it. Run all four against every output. The pass rate is the share that clears all four.
The shape of the failures tells you where the system breaks. A high visual score sitting next to a low claim score means the model paints well and overreaches on copy, and the rubric points straight at the seam. This diagnostic power is why compliance belongs inside a connected engine rather than a generator bolted on at the end. The engine can check output against the brand because it owns the brand context that produced the output. A standalone generator has no rule to enforce, only text to emit.
Reading the rate
Two numbers from the same axis behave very differently. A high visual pass rate is cheap; palette and logo are easy to lock. The claim axis is where benchmarks earn their keep, because that is where a fluent model invents a statistic or a promise the brand cannot stand behind. Weight your reporting accordingly. A high overall rate that hides a weak claim axis is not a passing grade. It is a flag telling you exactly which rule to harden before the next cycle.
The rubric travels across engines. W for Woman hit 86% image-gen success on virtual try-on through Charp.ai personalisation, a visual-axis number from a different brand and a different metric. The compliance question underneath is identical every time: measure the hit rate, name the weak axis, raise it.
Brand compliance and the cited answer
The benchmark reaches past creative. When a brand becomes the answer an AI model returns to a buyer, compliance is whether that answer is accurate and on-message. That is the remit of Echo IQ: being the cited, on-brand answer in the systems people now query before they search. The rubric holds. Does the model represent the brand the way the brand would represent itself. Score it, report it, raise it.
A brand that can score its compliance can improve it on a schedule. The next benchmark worth publishing is the one that tracks the rate climbing, cycle over cycle. See how the same discipline reads inside a live build across the case work.
Published 2026-06-20 · Whilter.AI