We investigate the validity of internal methods to estimate the uncertainty of the galaxy two-point correlation function. We consider the jackknife and bootstrap methods, which are based on re-sampling sub-regions of the original data. These are cheap computationally, and do not depend on the accuracy of external simulations. We test the different methods over a large range of scales using a set of 160 mock catalogues from the LasDamas set of simulations. Our results show that the standard bootstrap method significantly overestimates the true uncertainty at all scales. We try two possible generalisations of the bootstrap, but find them not to be robust. Regarding jackknife, we obtain that this method provides an unbiased estimation of the error at small and intermediate scales, up to ∼ 40 h−1, Mpc. At larger scales, it typically overestimates the error by a ∼13%.