Estimates of gene flow are commonly based on inferences of landscape resistance in ecological and evolutionary research and they frequently inform decision-making processes in conservation management. It is therefore imperative that inferences of landscape resistance are robust across approaches and reflect real-world gene flow instead of methodological artefacts. We used three empirical datasets that adopted individual-based sampling schemes and varied in scale (35-25,000 km squared) and total number of samples (184-790). These datasets comprise the wild boar, Sus scrofa, the red fox, Vulpes vulpes and the common wall lizard, Podarcis muralis. With a focus on comparing 160 different metrics to quantify individual-based pairwise genetic distances, we made use of a machine-learning algorithm implemented in RESISTANCEGA to optimally fit resistances of landscape factors to genetic distance metrics. Employed for nine landscape factors this resulted in 4,320 unique combinations of dataset, landscape factor and genetic distance metric, which provides the basis for quantifying uncertainty in inferences of landscape resistance. Our results demonstrate that there are clear differences in Akaike information criteria (AICc)-based model support and marginal R-squared-based model fit between different genetic distance metrics. Metrics based on between 1-10 axes of eigenvector-based multivariate analyses (Factorial correspondence analysis, FCA; Principal component analysis, PCA) outperformed more widely used metrics, including the proportion of shared alleles (DPS), with AICc and marginal R-squared values often an order of magnitude greater in the former. Across datasets, almost all categorical landscape features were inferred to either facilitate or impede gene flow depending on the choice of genetic distance metric. The directionality of the inferred resistance was largely consistent when considering metrics based on between 1-10 FCA/PCA axes. Inferences of landscape genetic resistance need to be corroborated using calculations of multiple individual-based pairwise genetic distance metrics. Our results call for the adoption of eigenvector-based quantifications of pairwise genetic distances. Specifically, a preliminary step of analysis should be incorporated, which explores model ranks across genetic distance metrics derived from FCA and PCA, and, contrary to findings of a simulation study, we demonstrate that it suffices to quantify these distances spanning the first ten axes only.