Testing whether structured brand guidance improves AI output
This anonymised case study examines a question many global brand teams are asking: does converting brand guidance into machine-readable data materially improve what AI produces? For a global professional services organisation, we designed a controlled evaluation comparing unguided generation, raw website guidance, a structured brand map, and the same map delivered through task-specific skills. The purpose was not to prove that AI could write plausible copy. It was to measure whether AI could follow the organisation’s rules, avoid unsupported claims, explain its decisions, and produce work a brand team could review with confidence.
The result was decisive. Without brand guidance, none of the evaluated pieces was ready for first-pass sign-off. With structured guidance, rule compliance rose to nearly nine in ten, invented facts disappeared, and reviewers judged nearly nine in ten pieces ready for first-pass approval. The skills condition achieved the strongest sign-off result and made none of the deliberately planted mistakes.
The challenge
The organisation had extensive brand guidance, but much of its meaning depended on relationships between rules, contextual conditions, examples, and content embedded inside images. Raw page text could give an AI model more words, but it could not reliably show which rule applied to a task or explain why a decision had been made.
The brand team needed evidence that a structured approach would improve more than tone. It had to improve compliance, factual discipline, task relevance, and the traceability of each judgement.
The evaluation design
We gave the same AI model four realistic briefs:
- a video script;
- an event plan;
- a social carousel; and
- a proposal.
Each brief was completed twice under four controlled conditions:
- no brand guidance;
- raw pages taken from the brand website;
- the structured brand map supplied as data; and
- the structured brand map delivered through task-specific skills.
Two independent reviewers scored every output blind. They did not know which condition had produced the work. Their answer key was built from the organisation’s original brand guidance, allowing every output to be tested against the same underlying rules.
Structuring the brand knowledge
We converted guidance written for human reading into a governed brand map. Each rule retained its source, conditions, relationships, and supporting evidence. AI could retrieve the guidance relevant to the task instead of loading large volumes of undifferentiated content.
The detailed work covered Tone of Voice, Photography Outside Our Library, Reinvention, and Video. Alongside structured Markdown and JSON outputs, we designed three cross-map skills for policy questions, content creation, and finished-work review.
What the evaluation found
Unguided AI followed only around a third of the relevant brand rules, and no output was considered ready for sign-off. Raw website pages improved plausibility, but important decisions remained unsupported and could not consistently be connected to the rule behind them.
Structured guidance more than doubled rule compliance. Nearly nine in ten rules were followed, and nearly nine in ten pieces were judged ready for first-pass sign-off. Only the structured conditions eliminated invented facts entirely. Unsupported specifics fell by more than half compared with the raw-page condition.
The skills condition performed best. It loaded only the guidance required for each brief, achieved the highest sign-off rate, and made none of the planted mistakes. Structure and task-specific retrieval worked together: the map made the knowledge explicit, while each skill applied the right part at the right moment.
Traceability changed the review process
In a separate evaluation across the four guidance families, we measured whether AI could connect decisions to the relevant rules and evidence. Scraped page text rarely created those connections. Structured guidance connected almost every decision to its supporting rule.
Tone of Voice showed why this matters. Many rules were communicated through images, so text scraping alone could not capture the full guidance. Once represented explicitly, reviewers could verify a judgement against its source in seconds instead of searching the brand site or relying on memory.
Outcome
The evaluation gave the organisation evidence that machine-readable guidance improves both AI output and human oversight. It established a measurable route from static guidelines to controlled creation and review, with decisions linked to the rule, condition, and source behind them.
The brand map remains a working asset for review and approval by the organisation’s brand team. The wider guidance has also been structured into themed skill sets spanning areas such as colour, typography, co-branding, and recruitment, creating a foundation for future governed generation and evaluation.
What this case study demonstrates
AI does not become brand-safe simply because it can access a guideline page. The improvement comes when brand knowledge is structured, scoped to the task, and connected to evidence. That turns brand judgement from an opaque model response into a decision a human reviewer can verify.