Abstract The spatial variability of topsoil water content (SWC) is often expressed through the relationship between its spatial mean 〈θ〉 and standard deviation σθ. The present study tests the concept that a reasonably performing land surface model (LSM) should be able to produce σθ–〈θ〉 data pairs that fall into a polygon, spanned by the cloud of observed data and two anchor points: σθ at the permanent wilting point σθ–〈θwp〉 and σθ at saturation σθ–〈θs〉. A state-of-the-art LSM, Noah-MP, was driven by atmospheric forcing data obtained from eddy covariance field measurements in two regions of southwestern Germany, Kraichgau (KR) and Swabian Alb (SA). KR is characterized with deep loess soils, whereas the soils in SA are shallow, clayey, and stony. The simulations series were compared with SWC data from soil moisture networks operating in the two study regions. The results demonstrate that Noah-MP matches temporal 〈θ〉 dynamics fairly well in KR, but performs poorly in SA. The best match is achieved with the van Genuchten–Mualem representation of soil hydraulic functions and site-specific rainfall, soil texture, green vegetation fraction (GVF) and leaf area index (LAI) input data. Nevertheless, most of the simulated σθ–〈θ〉 pairs are located outside the envelope of measurements and below the lower bound, which shows that the model smooths spatial SWC variability. This can be mainly attributed to missing topography and terrain information and inadequate representation of spatial variability of soil texture and hydraulic parameters, as well as the model assumption of a uniform root distribution.