Cloud radar is an invaluable tool for understanding complex microphysical processes unfolding in clouds. It operates at mm wavelengths where many, but not all, clouds are optically thin, thus justifying the single-scattering approximation in the 'radar equation' used to predict the radar signal. This equation is “inverted” to obtain cloud properties of interest from radar observations. More sophisticated predictive models for cloud radar signals account for multiple scattering, an inevitable consequence in nature of any scattering. However, they assume the cloud is made of layers, each horizontally uniform. In reality, they are not: clouds are complex 3D objects. Assuming otherwise biases retrieved cloud properties. To assess signal prediction error caused by the horizontal uniformity assumption, we developed a radiative transfer model to predict realistic radar signals accounting fully for 3D spatial variability that spans scales from large cloud systems (~100 km) down to 0.1 km, that is, far less than the CloudSat radar footprint (~1.5 km). As expected, single scattering is quickly overwhelmed by multiple scattering in the observed radar-pulse returns. We also find that “1D” operational models used in retrievals, even accounting for multiple scattering, fall short of correct signal prediction in proportions that depend on location and depth.