The adoption of Augmented Reality (AR) in the industry is in early stages, mainly due to technological and organizational limitations. This research work, carried out in a manufacturing factory, aims at providing an effective AR training method for manual assembly, adapted for industrial context. We define the 2W1H (What, Where, How) principle to formalize the description of any manual assembly operation in AR, independently on its type or complexity. Further, we propose a head-mounted display (HMD)-based method for conveying the manual assembly information, which relies on low-cost visual assets - i.e. text, image, video and predefined auxiliary content. We evaluate the effectiveness and usability of our proposal by conducting a field experiment with 30 participants. Additionally, we comparatively evaluate two sets of AR instructions, low-cost vs. CAD-based, to identify benefits of conveying assembly information by using CAD models. Our objective evaluation indicates that (i) manual assembly expertise can be effectively delivered by using spatially registered low-cost visual assets and that (ii) CAD-based instructions lead to faster assembly times, but persuade lower user attentiveness, eventually leading to higher error rates. Finally, by considering the diminishing utility of the AR instructions over three assembly cycles, we question the worthiness of authoring CAD-based AR instructions for similar industrial scenarios.