There is a long-standing tradition in the fields of system theory and sociophysics of approximating the dynamics of complex social networked systems with mean-field descriptions. Typically, such descriptions are derived under simplifying assumptions about network structure, with the most widely used simplification being that the system is fully mixed, and the corresponding network is a complete graph. Generally, more advanced descriptions that stem from more realistic network specifications lead to a higher level of agreement between direct simulations and mean-field solutions. In this paper, we consider two mean-field descriptions of an opinion formation model. One builds upon the complete mixing hypothesis, while the other is based on the assumption that the underlying social network is derived from a stochastic block model. Our finding is that, under specific conditions, the trajectories of the system obtained by means of direct simulations lie somewhere between the two approximations. We report that this is the case at the intersection of the following conditions: (i) sparse and extremely modular network topologies, (ii) bounded confidence opinion dynamics, and (iii) a personalization algorithm that fosters contacts between individuals with diverse opinions.