This paper aims to investigate how stability varies with the approach used in estimating school performance in a large sample of English primary schools. The results show that (a) raw performance is considerably more stable than adjusted performance, which in turn is slightly more stable than growth model estimates; (b) schools' performance indicators are less dependent on the schools' social composition score when scores are adjusted, and in particular for growth scores; (c) in the multilevel growth models (MGM), 74% of the variance in student rates of growth is attributable to schools; (d) the trajectories of individual students within a school are highly convergent, as if they were all being brought towards a common, stable target. Some implications for using these indicators for evaluating and regulating school improvement are discussed.