In this article, we explore the impact of outfits on AI-based Pedestrian Dead Reckoning (PDR) with a pocket-worn inertial sensor. This PDR mode faces significant variability due to the countless choices of outfits available. We observe significant variations in the inertial signals captured by a pocket-worn device, which are highly influenced by the outfit being worn. To address this, we propose a 2-category classification of outfits as tight or loose, based on their impact on the inertial signals. Notably, AI models trained on tight outfits exhibit poor generalization with loose outfits and vice versa. We highlight this phenomenon by implementing a data-low-cost PDR algorithm based on Support Vector Regression (SVR) and assess its performance on two healthy volunteers and a senior and blind volunteer wearing tight and loose outfits, on real-life situation test tracks spanning approximately 200 to 400 meters.