Context-Aware Criteria Generation with VLMs for Advertisement Ranking under Data Scarcity

Vision-Language Models (VLMs) perform strongly on generic multimodal reasoning tasks, yet real-world business decisions require reasoning that depends on the specific business objective under very limited data and labels. We introduce a new task, brand-specific advertisement ranking, which aims to rank ads for a target brand under these constraints. This task requires capturing what makes an ad effective for that particular brand, rather than relying on generic visual-textual cues or large-scale supervision. To tackle this task, we propose ADvisor, a novel approach that derives explicit brand-aware decision criteria with careful use of VLMs. ADvisor also augments the limited brand-specific context with ads from similar brands, and uses the generated criteria for reflection-driven scoring to rank ads. Experiments on real-world advertising data from 10 brands, which include performance labels for evaluation, collected from actual ad campaigns, show that ADvisor outperforms the strongest baseline by up to 7.2% in ranking performance. ADvisor also performs strongly in online A/B testing, even when compared with human experts. In addition, both qualitative and quantitative analyses confirm that the generated criteria capture effective brand-specific evaluation standards rather than generic ones. Our code is available at https://github.com/K-Kyungho/ADvisor.