The eddy covariance technique, which is used in the determination of net ecosystem CO2 exchange (NEE), is subject to significant errors when advection that carries CO2 in the mean flow is ignored. We measured horizontal and vertical advective CO2 fluxes at the Niwot Ridge AmeriFlux site (Colorado, USA) using a measurement approach consisting of multiple towers. We observed relatively high rates of both horizontal (F(hadv)) and vertical (F(vadv)) advective fluxes at low surface friction velocities (u(*)) which were associated with downslope katabatic flows. We observed that F(hadv) was confined to a relatively thin layer (0-6 m thick) of subcanopy air that flowed beneath the eddy covariance sensors principally at night, carrying with it respired CO2 from the soil and lower parts of the canopy. The observed F(vadv) came from above the canopy and was presumably due to the convergence of drainage flows at the tower site. The magnitudes of both F(hadv) and F(vadv) were similar, of opposite sign, and increased with decreasing u(*), meaning that they most affected estimates of the total CO2 flux on calm nights with low wind speeds. The mathematical sign, temporal variation and dependence on u(*) of both F(hadv) and F(vadv) were determined by the unique terrain of the Niwot Ridge site. Therefore, the patterns we observed may not be broadly applicable to other sites. We evaluated the influence of advection on the cumulative annual and monthly estimates of the total CO2 flux (F(c)), which is often used as an estimate of NEE, over six years using the dependence of F(hadv) and F(vadv) on u(*). When the sum of F(hadv) and F(vadv) was used to correct monthly F(c), we observed values that were different from the monthly F(c) calculated using the traditional u(*)-filter correction by--16 to 20 g C x m(-2) x mo(-1); the mean percentage difference in monthly Fc for these two methods over the six-year period was 10%. When the sum of F(hadv) and F(vadv) was used to correct annual Fc, we observed a 65% difference compared to the traditional u(*)-filter approach. Thus, the errors to the local CO2 budget, when F(hadv) and F(vadv) are ignored, can become large when compounded in cumulative fashion over long time intervals. We conclude that the "micrometeorological" (using observations of F(hadv) and F(vadv)) and "biological" (using the u(*) filter and temperature vs. F(c) relationship) corrections differ on the basis of fundamental mechanistic grounds. The micrometeorological correction is based on aerodynamic mechanisms and shows no correlation to drivers of biological activity. Conversely, the biological correction is based on climatic responses of organisms and has no physical connection to aerodynamic processes. In those cases where they impose corrections of similar magnitude on the cumulative F(c) sum, the result is due to a serendipitous similarity in scale but has no clear mechanistic explanation.