User sequence modeling is widely used in recommendation and advertising. Sequence modules improve click and conversion prediction while satisfying auction-time serving constraints in large-scale sponsored ranking systems. We present WaySeq, Wayfair's first production framework for sequence-aware sponsored-products ranking, and use it to study which sequence modeling choices are most effective under these constraints. Starting from a lightweight transformer backbone, we evaluate sequence coverage, embedding representation, sequence summarization, tokenization, and attention structure. Our results show that the largest gains come from broader behavioral coverage, stronger product embeddings, and candidate-aware sequence summarization, while heavier behavior-aware attention variants add serving cost without consistent ranking-quality gains. The final production model delivers consistent online gains across engagement and conversion-oriented metrics. These findings suggest that, in industrial sponsored ranking, the interface between user history and the ad candidate, together with representation quality, is more important than simply increasing architectural complexity.