Stable isotope analysis has become one of the most widely used techniques in ecology. However, uncertainties about the effects of sample preservation and pre-treatment on the ecological interpretation of stable isotope data and especially on Bayesian stable isotope mixing models remain. Here, Bayesian mixing models were used to study how three different preservation methods (drying, freezing, formalin) and two pre-treatments (acidification, lipid removal) affect the estimation of diet composition for two benthic invertebrate species (<i>Limecola balthica, Crangon crangon</i>). Furthermore, commonly used mathematical lipid normalization and formalin correction were applied to check if they improve the model results. Preservation effects were strong on model outcomes for frozen as well as formalin preserved <i>L. balthica</i> samples, but not for <i>C. crangon</i>. Pre-treatment effects varied with species and preservation method and neither lipid normalization nor mathematical formalin correction consistently resulted in improved model outcomes. Our analysis highlights that particularly small changes in δ15N introduced by different preservation and pre-treatments display a so far unrecognized source of error in stable isotope studies. We conclude that mathematical correction of stable isotopes data should be avoided for Bayesian mixing models and that previously unaddressed effects of sample preservation (especially those arising from preservation by freezing) have potentially biased our understanding of the utilization of organic matter in aquatic food webs.