Advertising
Finding the audiences a creative actually works for.
Multimodal models that learn what is in each ad creative and match it to the audience segments most likely to respond, uncovering targeting overlaps people had missed.
- Engagement
- Team track record
- Context
- Global beverage brand, via a global advertising group
- Capabilities
- Multimodal learning · Audience modelling · ML engineering
- ad revenue uplift
- +10%
The problem
Large brands produce hundreds of creatives and target dozens of audience segments. Which creative goes to which audience was decided largely by intuition, and valuable combinations went unnoticed.
Our approach
- Multimodal creative embeddings captured what each ad shows and says: imagery, text, tone and format, in one representation.
- Cross-modal attention learned how those creative features relate to the behaviour of high-value audience segments.
- Ranked matches surfaced non-obvious creative-audience pairs for media teams to act on.
The result
The recommendations drove a 10% uplift in ad revenue, largely from targeting overlaps that weren’t visible in the existing reports.
Team track record: work led by members of our team in previous roles, described without naming clients or employers. Demos are illustrative, built on synthetic data and fictional brands.