A new observational approach is presented to approximate the uncertainty (scatter or error variance) in 1-h averaged turbulence fluxes from eddy-covariance measurements. The uncertainty includes potential contributions from instrument problems, heterogeneity and non-stationarity in addition to classical random sampling error. The daytime relative flux uncertainty (RFE) is half as large (20%) at a simple maize site compared to two more complex forest sites (40%) for all scalars possibly due to the more homogeneous vegetation, flat terrain and especially the lower measurement height. The RFE is approximately the same for day and night periods for all scalars at the two mostly homogeneous sites (pine forest and maize field) except for latent heat over the forest, where the RFE doubles at night. Modest surface heterogeneity at the other forest site for nocturnal flux footprints approximately doubles the RFE compared to daytime conditions for all scalar fluxes. Compositing by atmospheric stability (instead of time of day) reveals a sharp increase in the RFE for the most strongly stable conditions. A theoretical prediction for the pure random sampling error based on the flux integral timescale is smaller by a factor of two compared to the observed variability.