Contextual advertising works by aligning ads with the media environment they appear in. Our research introduces a framework for identifying key ad and context features, using machine learning to predict ad and brand attention.

We leverage a Multimodal Large Language Model to analyze high-level topics and an XGBoost model for prediction, trained on eye-tracking data from 3,531 digital ads.

Our findings show that both XGBoost and ResNet50 accurately predict attention, with ResNet50 excelling in ad-focused attention and XGBoost leading in brand attention. Notably, for new, unseen ads, our theory-driven XGBoost model outperforms ResNet50, offering a more reliable approach for contextual ad targeting. Shapley values highlight which ad and context features drive engagement, reinforcing the power of AI-driven, theory-backed advertising strategies.