Natural history collections are alternative data sources to plot-based species inventories for analysing macroecological species turnover. Herbarium records sample diversity well at regional level and are taxonomically validated. However, they are ad hoc from a sampling perspective, generating spatial and taxonomic biases. The implications of biased sampling on beta diversity (β) estimation, and use of herbarium data to identify macroecological transitions, remain unexplored. We tested sampling influences by comparing herbarium data with systematically collected inventory data from the Mount Lofty-Flinders Ranges region of Australia. We calculated β within moving windows across bioclimatic gradients using metrics varying in sensitivity to richness differences (pairwise/multi-site Sørensen β Simpson β Harrison et al. β-2), and correlated β to species sampling and between herbarium and plot data. We tested whether generalised dissimilarity modelling (GDM) revealed the same compositional transitions in herbarium and plot data along environmental gradients. Sørensen, Simpson and multi-site Sørensen β had strong negative correlations with richness (indicating sampling bias) for herbarium data (Pearson's r=-0.85, -0.80, -0.81, respectively) but not plots (r=-0.27, -0.28, -0.11). Harrison et al. β-2 correlated poorly with richness (herbarium: r=-0.16; plots: r=-0.14) but herbarium and plot data were only weakly correlated (r=0.18). All other metrics correlated poorly (-0.03<r<0.16) between datasets, suggesting biases. GDMs differed in variable importance but revealed similar transition zones for key gradients. We conclude that untransformed herbarium data are unsuitable for detecting macroecological transitions because turnover is linearly related to sampling intensity and correlates poorly with systematic surveys. Herbarium data should be used cautiously for β, even with methods insensitive to richness differences. However, herbarium data can robustly reproduce transition zones when modelled along environment gradients. We recommend this approach for detecting macroecological transitions using natural history data in the absence of plot data.