Sediments contained in the river bed do not necessarily contribute to morphological change. The finest part of the sediment mixture often fills the pores between the larger grains and can be removed without causing a drop in bed level. The discrimination between pore-filling load and bed-structure load, therefore, is of practical importance for morphological predictions. In this study, a new method is proposed to estimate the cut-off grain size that forms the boundary between pore-filling load and bed-structure load. The method evaluates the pore structure of the river bed geometrically. Only detailed grain-size distributions of the river bed are required as input to the method. A preliminary validation shows that the calculated porosity and cut-off size values agree well with experimental data. Application of the new cut-off size method to the river Rhine demonstrates that the estimated cut-off size decreases in a downstream direction from about 2 to 0·05 mm, covariant with the downstream fining of bed sediments. Grain size fractions that are pore-filling load in the upstream part of the river thus gradually become bed-structure load in the downstream part. The estimated (mass) percentage of pore-filling load in the river bed ranges from 0% in areas with a unimodal river bed, to about 22% in reaches with a bimodal sand-gravel bed. The estimated bed porosity varies between 0·15 and 0·35, which is considerably less than the often-used standard value of 0·40. The predicted cut-off size between pore-filling load and bed-structure load (Dc,p) is fundamentally different from the cut-off size between wash-load and bed-material load (Dc,w), irrespective of the method used to determine Dc,p or Dc,w. Dc,w values are in the order of 10−1 mm and mainly dependent on the flow characteristics, whereas Dc,p values are generally much larger (about 100 mm in gravel-bed rivers) and dependent on the bed composition. Knowledge of Dc,w is important for the prediction of the total sediment transport in a river (including suspended fines that do not interact with the bed), whereas knowledge of Dc,p helps to improve morphological predictions, especially if spatial variations in Dc,p are taken into account. An alternative to using a spatially variable value of Dc,p in morphological models is to use a spatially variable bed porosity, which can also be predicted with the new method. In addition to the morphological benefits, the new method also has sedimentological applications. The possibility to determine quickly whether a sediment mixture is clast-supported or matrix-supported may help to better understand downstream fining trends, sediment entrainment thresholds and variations in hydraulic conductivity.