We investigate an extension to the probabilistic model of a wireless sensor network (WSN) in a recently proposed variational message passing (VMP) localization algorithm. This extension exploits network topology information to mitigate ambiguities in WSN localization schemes. We derive VMP schemes for direct (1-step) localization and 2-step localization where knowledge on the positions of 2-step neighbours is taken into account. We evaluate the performance of the algorithms in connected WSNs via Monte Carlo simulations. The results show that utilizing position information from sensors with which there is no direct connection improves the position estimates for sensors directly connected to few neighbour sensors. The most significant improvement is observed for sensors of degree two i.e. with two 1-step neighbours. However, further investigations considering more realistic WSNs with decentralized and (partly) centralized settings are required to obtain more conclusive results on the real performance gain achievable by exploiting information on absent connections.