A Machine-Readable Rulebook for Classifying Advertising
Media Focus classifies advertising against an extensive rulebook. Together, we made that rulebook machine-readable and built an AI that classifies new ads against it and explains every suggestion.
- Client
- Media Focus
- Status
- In production
“In Voviva, we found a partner who understood our requirements right away. The straightforward collaboration and the fast, professional delivery laid the ideal foundation for expanding and scaling the use of AI at Media Focus for the long term.”
Michel BächtigerCDO, Media FocusAt a glance
One rulebook for all
The rulebook is the single source of all classification rules, for the team and the AI.
Reasoned suggestions
The AI explains every classification and only chooses values that exist in the rulebook.
The team stays in control
The specialists maintain the rulebook themselves and correct where needed.
The challenge
Media Focus monitors the Swiss advertising market and classifies advertising systematically, for example by industry, product group and advertised products. The basis is an extensive rulebook that has grown over the years.
That rulebook was spread across separate documents (e.g. Word, PDF, PowerPoint). Machines could not read it, and much of what matters for a correct classification lived only in the team's experience. The classification itself was done manually.
Our solution
First, the rulebook was given a consistent structure. Every category now comprises a definition, typical terms and clear distinctions that both people and machines can understand. The team maintains the rulebook itself in a simple editor, without any code. Every change is saved and traceable, and the rulebook can be exported as a document at any time.
On that basis, a multi-stage AI pipeline classifies new ads, whether as an image, PDF or text. It first captures the content, brand and advertising message, then classifies the ad systematically, step by step in increasing detail. It can only choose values that exist in the rulebook, and it explains every suggestion.
The specialists review the suggestions and correct them where needed. An overview of corrections shows where the system is still off, and test sets built from existing classifications make quality measurable.