ABSTRACT We re-analyse data from the gravitational-wave event GW170817 and its host galaxy NGC 4993 to demonstrate the importance of accurate total and peculiar velocities when measuring the Hubble constant using this nearby standard siren. We show that a number of reasonable choices can be made to estimate the velocities for this event, but that systematic differences remain between these measurements depending on the data used. This leads to significant changes in the Hubble constant inferred from GW170817. We present Bayesian model averaging as one way to account for these differences, and obtain $H_{0}=66.8^{+13.4}_{-9.2}\, \mathrm{km\, s^{-1}\, Mpc^{-1}}$. Adding additional information on the viewing angle from high-resolution imaging of the radio counterpart refines this to $H_{0}=64.8^{+7.3}_{-7.2}\, \mathrm{km\, s^{-1}\, Mpc^{-1}}$. During this analysis, we also present an alternative Bayesian model for the posterior on H0 from standard sirens that works more closely with observed quantities from redshift and peculiar velocity surveys. Our results more accurately capture the true uncertainty on the total and peculiar velocities of NGC 4993 and show that exploring how well different data sets characterize galaxy groups and the velocity field in the local Universe could improve this measurement further. These considerations impact any low-redshift distance measurement, and the improvements we suggest here can also be applied to standard candles like Type Ia supernovae. GW170817 is particularly sensitive to peculiar velocity uncertainties because it is so close. For future standard siren measurements, the importance of this error will decrease as (i) we will measure more distant standard sirens and (ii) the random direction of peculiar velocities will average out with more detections.