The standard approach to flood frequency analysis (FFA) fits mathematical functions to sequences of historic flood data and extrapolates the tails of the distribution to estimate the magnitude and likelihood of extreme floods. Here, we identify the most exceptional floods in the United States as compared against other major floods at the same location, and evaluate how the flood-of-record (Qmax) influences FFA estimates. On average, floods-of-record are 20% larger by discharge than their second-place counterparts (Q2), and 212 gages (7.3%) have Qmax:Q2 ratios greater than two. There is no clear correspondence between the Qmax:Q2 ratio and median instantaneous discharge, and exceptional floods do not become less likely with time. Excluding Qmax from the FFA causes the median 100-year flood to decline by −10.5%, the 200-year flood by −11.8%, and the 500-year flood by −13.4%. Even when floods are modelled using a heavy tail distribution, the removal of Qmax yields significantly “lighter” tails and underestimates the risk of large floods. Despite the temporal extension of systematic hydrological observations in the United States, FFA is still sensitive to the presence of extreme events within the sample used to calculate the frequency curve.