Infiltrating river water carries the temperature signal of the river into the adjacent aquifer. While the diurnal temperature fluctuations are strongly dampened, the seasonal fluctuations are much less attenuated and can be followed into the aquifer over longer distances. In one-dimensional model with uniform properties, this signal is propagated with a retarded velocity, and its amplitude decreases exponentially with distance. Therefore, time shifts in seasonal temperature signals between rivers and groundwater observation points may be used to estimate infiltration rates and near-river groundwater velocities. As demonstrated in this study, however, the interpretation is nonunique under realistic conditions. We analyze a synthetic test case of a two-dimensional cross section perpendicular to a losing stream, accounting for multi-dimensional flow due to a partially penetrating channel, convective-conductive heat transport within the aquifer, and heat exchange with the underlying aquitard and the land surface. We compare different conceptual simplifications of the domain in order to elaborate on the importance of different system elements. We find that temperature propagation within the shallow aquifer can be highly influenced by conduction through the unsaturated zone and into the underlying aquitard. In contrast, regional groundwater recharge has no major effect on the simulated results. In our setup, multi-dimensionality of the flow field is important only close to the river. We conclude that over-simplistic analytical models can introduce substantial errors if vertical heat exchange at the aquifer boundaries is not accounted for. This has to be considered when using seasonal temperature fluctuations as a natural tracer for bank infiltration.