Since the first estimate of global CO2 emissions was published in 1894, important progress has been made in the development of estimation methods while the number of available datasets has grown. The existence of parallel efforts should lead to improved accuracy and understanding of emissions estimates, but there remains significant deviation between estimates and relatively poor understanding of the reasons for this. Here I describe the most important global emissions datasets available today and – by way of global, large-emitter, and case examples – quantitatively compare their estimates, exploring the reasons for differences. In many cases differences in emissions come down to differences in system boundaries: which emissions sources are included and which are omitted. With minimal work in harmonising these system boundaries across datasets, the range of estimates of global emissions drops to 5 %, and further work on harmonisation would likely result in an even lower range, without changing the data. Some potential errors were found, and some discrepancies remain unexplained, but it is shown to be inappropriate to conclude that uncertainty in emissions is high simply because estimates exhibit a wide range. While “true” emissions cannot be known, by comparing different datasets methodically, differences that result from system boundaries and allocation approaches can be highlighted and set aside to enable identification of true differences, and potential errors. This must be an important way forward in improving global datasets of CO2 emissions. Data used to generate Figs. 3–18 are available at https://doi.org/10.5281/zenodo.3687042 (Andrew, 2020).