Utilising a fleet of commercial airliners, MOZAIC/IAGOS provides atmospheric composition data on a regular basis that are widely used for modelling applications. Due to the specific operational context of the platforms, such observations are collected close to international airports and hence in an environment characterised by high anthropogenic emissions. This provides opportunities for assessing emission inventories of major metropolitan areas around the world, but also challenges in representing the observations in typical chemical transport models. We assess here the contribution of different sources of error to overall model–data mismatch using the example of MOZAIC/IAGOS carbon monoxide (CO) profiles collected over the European regional domain in a time window of 5 yr (2006–2011). The different sources of error addressed in the present study are: 1) mismatch in modelled and observed mixed layer height; 2) bias in emission fluxes and 3) spatial representation error (related to unresolved spatial variations in emissions). The modelling framework combines a regional Lagrangian transport model (STILT) with EDGARv4.3 emission inventory and lateral boundary conditions from the MACC reanalysis. The representation error was derived by coupling STILT with emission fluxes aggregated to different spatial resolutions. We also use the MACC reanalysis to assess uncertainty related to uncertainty sources 2) and 3). We treat the random and the bias components of the uncertainty separately and found that 1) and 3) have a comparable impact on the random component for both models, while 2) is far less important. On the other hand, the bias component shows comparable impacts from each source of uncertainty, despite both models being affected by a low bias of a factor of 2–2.5 in the emission fluxes. In addition, we suggested methods to correct for biases in emission fluxes and in mixing heights. Lastly, the evaluation of the spatial representation error against model–data mismatch between MOZAIC/IAGOS observations and the MACC reanalysis revealed that the representation error accounts for roughly 15–20% of the model–data mismatch uncertainty.