We generate mock galaxy catalogues for a grid of different cosmologies, using rescaled N-body simulations in tandem with a semi-analytic model run using consistent parameters. Because we predict the galaxy bias, rather than fitting it as a nuisance parameter, we obtain an almost pure constraint on σ8 by comparing the projected two-point correlation function we obtain to that from the Sloan Digital Sky Survey (SDSS). A systematic error arises because different semi-analytic modelling assumptions allow us to fit the r-band luminosity function equally well. Combining our estimate of the error from this source with the statistical error, we find σ8= 0.97 ± 0.06. We obtain consistent results if we use galaxy samples with a different magnitude threshold, or if we select galaxies by bJ-band rather than r-band luminosity and compare to data from the 2dF Galaxy Redshift Survey (2dFGRS). Our estimate for σ8 is higher than that obtained for other analyses of galaxy data alone, and we attempt to find the source of this difference. We note that in any case, galaxy clustering data provide a very stringent constraint on galaxy formation models.