We address the issue of how to compare atmospheric carbon dioxide (CO2) observations from the National Oceanic and Atmospheric Administration/Climate Monitoring and Diagnostics Laboratory (NOAA/CMDL) air sampling network to model output in a consistent method for model evaluation and inverse problems. Two steps are needed to sample the model using the same methodology that is applied to the flask samples: (1) ensure that the wind direction is one that is deemed acceptable in the field and (2) reject outliers. Since a clean air sector has only been determined for four sites, we compute a statistical analysis of the wind direction when samples are collected at 38 NOAA/CMDL CO2 collection sites operational from 1993--1995. We compare this to the wind directions from the European Centre for Medium-Range Weather Forecasts (ECMWF) assimilated wind fields for the same 3 year time period. From this analysis, we deduce which of the sites are selecting for wind direction. We then analyze model output to compare the effects on monthly mean concentration from this two-step methodology. We find that at the majority of sites the removal of outliers is particularly important. Finally, we explore the impact of sampling frequency in the model. At most sites, the sampling frequency does impact the model results. We conclude that it is best to use high-frequency model output rather than sampling with the same low-frequency as the flask network. The bias resulting from low-frequency sampling is potentially large, and if done in both the model and the observations, the comparison error is exacerbated. We estimate this bias error for use in error estimation and inverse problems.¿ 1997 American Geophysical Union