Brand guidance has to become machine data before AI can use it reliably
This anonymised case study shows how a complex digital brand system was transformed into machine-readable data without flattening its meaning. The source guidance combined page text, designed examples, photography, diagrams, and rules communicated visually. A conventional website scrape could collect visible words, but it could not reliably capture what an image demonstrated, how one rule qualified another, or which condition made a rule relevant to a particular task.
We converted that mixed source material into explicit, governed brand data. Rules were separated from page layouts and represented with their conditions, relationships, examples, sources, and intended uses. The resulting structured Markdown and JSON created a dependable knowledge layer for policy questions, content creation, finished-work review, API delivery, and AI evaluation.
The challenge
Brand websites are made for people. Meaning is often distributed between copy, visual examples, captions, layouts, and implied relationships. That creates several problems for AI:
- text extraction misses guidance communicated through images;
- isolated passages lose the conditions governing their use;
- examples can be mistaken for universal rules;
- related requirements are retrieved separately; and
- generated answers cannot explain where a decision came from.
Tone of Voice was a particularly important example. Many of its practical rules lived inside designed visual material, so page text alone could not represent the complete guidance.
The data preparation method
We reviewed each guidance family as a knowledge system rather than a collection of pages. The work identified:
- the individual rule being expressed;
- the condition determining when it applies;
- relationships to other rules and exceptions;
- positive and negative examples;
- source location and evidence;
- ownership and approval status; and
- the tasks and skills allowed to use it.
Where guidance was communicated visually, the meaning was extracted and represented explicitly rather than leaving an AI model to infer it repeatedly from an image.
Building the machine-readable layer
The structured outputs were produced in Markdown and JSON. Human-readable Markdown supported review by brand owners, while JSON provided predictable fields for retrieval, APIs, evaluation, and agent skills.
The model preserved provenance. A retrieved rule could carry the source behind it, allowing a generated decision or correction to point back to the approved guidance. This made the data useful not only for generation, but also for assurance and audit.
Testing whether the structure retained meaning
We evaluated how well AI connected decisions to related rules and evidence across Tone of Voice, Photography Outside Our Library, Reinvention, and Video.
With scraped page text, the system rarely connected decisions to the rules behind them. With structured guidance, it connected almost all evaluated decisions to the relevant rules and evidence across every family. The improvement was particularly clear where page text had omitted meaning carried by visual examples.
Outcome
The organisation gained a machine-readable brand-data layer that preserved more than the words on its website. Rules, context, relationships, examples, and evidence became individually retrievable and reviewable.
This created the foundation for fifteen themed skill sets and three cross-map operating skills covering policy answers, content creation, and finished-work review. The structured assets remain subject to approval by the organisation’s brand team before production release.
What this case study demonstrates
Better AI output begins with better machine data. When brand meaning is represented explicitly, agents no longer have to reconstruct policy from page layouts or infer the lesson inside an image. They can retrieve the approved rule, apply it under the correct condition, and show the evidence behind the result.