Transverse-momentum-dependent (TMD) observables play a central role in many collider experiments and provide an important testing ground for precision Standard Model predictions. TMD factorization, which exploits the hierarchy between the measured transverse momentum qT and the hard scale Q, provides a systematic framework for separating perturbative and non-perturbative physics while also enabling the resummation of large logarithms. Although leading-power TMD factorization is by now well understood, the increasing precision of LHC measurements and the future Electron-Ion Collider make a corresponding understanding at next-to-leading power increasingly important. Several frameworks for next-to-leading-power TMD factorization have been considered in the literature, among which are: (1) the TMD operator product expansion, (2) SCET in the label formalism, and (3) SCET in the position-space formalism using the background-field approach. While these frameworks address the same physical problem, they appear to yield different results in the soft sector. In particular, the two SCET formulations appear to predict different soft contributions, while the TMD operator product expansion contains no explicit soft degrees of freedom at all. In this talk, I will present next-to-leading-power TMD factorization in the position-space formulation of SCET and discuss the apparent tensions between the different formalisms regarding the role of sub-leading soft contributions.