Organic vs. Synthetic Attention: Evaluating the Utility of Eye Tracking Augmentation in Recommender Systems

Eye-Tracking (ET) has recently gained popularity in Recommender Systems (RS) as a source of implicit feedback, effectively linking user attention to product saliency. However, leveraging user visual behaviour for recommendations does not scale due to the large volumes of ET data required to train RS models and the need for specialized equipment. This work investigates the extent to which state-of-the-art ET generative models can effectively mimic user information foraging behaviour in query-driven, search-recommendation tasks, as well as augment organic datasets with synthetic data for use in ET-based RS. We benchmark saliency and scan-path generators under task context, evaluating aggregate and individual saliency metrics and testing whether synthetic ET preserves category-sensitive gaze differences observed in humans. Our findings demonstrate that synthetic ET reproduces aggregate attention but fails to capture the search dynamics observed in organic data. Individual-level predictive accuracy remains low, with moderate improvements from fine-tuning and leave-one-out training, indicating data limitations.