In the presence of measurement noise, the quality of fast randomized low-rank approximations deteriorates with noise level. Under suitable assumptions, this phenomenon can be quantified (as function of the noise level, sketch size and underlying matrix rank) and mitigated.