Pollen productivity estimates (PPEs) are a key parameter for quantitative land-cover reconstructions from pollen data. PPEs are commonly estimated using modern pollen-vegetation data sets and the extended R-value (ERV) model. Prominent discrepancies in the existing studies question the reliability of the approach. We here propose an implementation of the ERV model in the R environment for statistical computing, which allows for simplified application and testing. Using simulated pollen-vegetation data sets, we explore sensitivity of ERV application to (1) number of sites, (2) vegetation structure, (3) basin size, (4) noise in the data, and (5) dispersal model selection. The simulations show that noise in the (pollen) data and dispersal model selection are critical factors in ERV application. Pollen count errors imply prominent PPE errors mainly for taxa with low counts, usually low pollen producers. Applied with an unsuited dispersal model, ERV tends to produce wrong PPEs for additional taxa. In a comparison of the still widely applied Prentice model and a Lagrangian stochastic model (LSM), errors are highest for taxa with high and low fall speed of pollen. The errors reflect the too high influence of fall speed in the Prentice model. ERV studies often use local scale pollen data from for example, moss polsters. Describing pollen dispersal on his local scale is particularly complex due to a range of disturbing factors, including differential release height. Considering the importance of the dispersal model in the approach, and the very large uncertainties in dispersal on short distance, we advise to carry out ERV studies with pollen data from open areas or basins that lack local pollen deposition of the taxa of interest.