Non-Gaussianity in the primordial fluctuations that seeded structure formation produces a signal in the galaxy power spectrum on very large scales. This signal contains vital information about the primordial Universe, but it is very challenging to extract, because of cosmic variance and large-scale systematics-especially after the Planck experiment has already ruled out a large amplitude for the signal. Whilst cosmic variance and experimental systematics can be alleviated by the multitracer method, we here address another systematic-introduced by not using the correct relativistic analysis of the power spectrum on very large scales. In order to reduce the errors on f NL , we need to include measurements on the largest possible scales. Failure to include the relativistic effects on these scales can introduce significant bias in the best-fit value of f NL from future galaxy surveys.