We test the critical point concept for earthquakes in terms of the spatial correlation length. A system near a critical point is associated with a diverging correlation length following a power law time-to-failure relation. We estimate the correlation length directly from an earthquake catalog using single-link cluster analysis. Therefore we assume that the distribution of moderate earthquakes reflects the state of the regional stress field. The parameters of the analysis are determined by an optimization procedure, and the results are tested against a Poisson process with realistic distributions of epicenters, magnitudes, and aftershocks. A systematic analysis of all earthquakes with M≥6.5 in California since 1952 is conducted. In fact, we observe growing correlation lengths in most cases. The null hypothesis that this behavior can be found in random data is rejected with a confidence level of more than 99%. Furthermore, we find a scaling relation log R~0.7M (log⟨&xgr;max⟩~0.5M), between the mainshock magnitude M and the critical region R (the correlation length ⟨&xgr;max⟩ before the mainshock), which is in good agreement with theoretical values. ¿ 2001 American Geophysical Union