Operational retrievals of tropospheric trace gases from space-borne spectrometers are based on one-dimensional radiative transfer models. To minimize cloud effects, trace gas retrievals generally implement a simple cloud model based on radiometric cloud fraction estimates and photon path length corrections. The latter relies on measurements of the oxygen collision pair (O2–O2) absorption at 477 nm or on the oxygen A-band around 760 nm to determine an effective cloud height. In reality however, the impact of clouds is much more complex, involving unresolved sub-pixel clouds, scattering of clouds in neighbouring pixels, and cloud shadow effects, such that unresolved three-dimensional effects due to clouds may introduce significant biases in trace gas retrievals. Although clouds have significant effects on trace gas retrievals, the current cloud correction schemes are based on a simple cloud model, and the retrieved cloud parameters must be interpreted as effective values. Consequently, it is difficult to assess the accuracy of the cloud correction only based on analysis of the accuracy of the cloud retrievals, and this study focuses solely on the impact of the 3D cloud structures on the trace gas retrievals. In order to quantify this impact, we study NO2 as a trace gas example and apply standard retrieval methods including approximate cloud corrections to synthetic data generated by the state-of-the-art three-dimensional Monte Carlo radiative transfer model MYSTIC. A sensitivity study is performed for simulations including a box cloud, and the dependency on various parameters is investigated. The most significant bias is found for cloud shadow effects under polluted conditions. Biases depend strongly on cloud shadow fraction, NO2 profile, cloud optical thickness, solar zenith angle, and surface albedo. Several approaches to correct NO2 retrievals under cloud shadow conditions are explored. We find that air mass factors calculated using fitted surface albedo or corrected using the O2–O2 slant column density can partly mitigate cloud shadow effects. However, these approaches are limited to cloud-free pixels affected by surrounding clouds. A parameterization approach is presented based on relationships derived from the sensitivity study. This allows measurements to be identified for which the standard NO2 retrieval produces a significant bias and therefore provides a way to improve the current data flagging approach.