Abstract We investigate the expected cosmological constraints from a combination of cosmic shear and large-scale galaxy clustering using realistic photometric redshift distributions. Introducing a systematic bias in the lensing distributions (of 0.05 in redshift) produces a >2σ bias in the recovered matter power spectrum amplitude and dark energy equation of state for preliminary Stage III surveys. We demonstrate that cosmological error can be largely removed by marginalizing over biases in the assumed weak-lensing redshift distributions. Furthermore, the cosmological constraining power is retained despite removing much of the information on the lensing redshift biases. This finding relies upon high-quality redshift estimates for the clustering sample, but does not require spectroscopy. All galaxies in this analysis can thus be assumed to come from a single photometric survey. We show that this internal constraint on redshift biases arises from complementary degeneracy directions between cosmic shear and the combination of galaxy clustering and shear–density cross-correlations. Finally we examine a case where the assumed redshift distributions differ from the truth by more than a simple uniform bias. We find that the effectiveness of this self-calibration method will depend on the survey details and the nature of the uncertainties on the estimated redshift distributions.