Abstract In order to better understand the preparation zone of the predicted Parkfield earthquake, a detailed study of the seismicity at middle Mountain in the Parkfield, California, area was made using 71 digitally recorded earthquakes that located within, or close to, the Middle Mountain alert box. These earthquakes were retimed on an interactive graphics system. Based on these new arrival times, new station corrections were developed; however the data did not support changing the velocity model developed from refraction and 1966 aftershock data. The process of retiming the earthquakes and using the new station corrections reduced the rms travel-time residuals by 70 per cent to 0.025 sec, halved the location errors, and clustered the earthquakes closer to the surface trace of the San Andreas fault. The seismicity can be approximated by a plane on the scale of several kilometers, but at finer scales two clusters were discovered that show demonstrable width to the seismogenic zone. Previous workers had proposed a 5° bend in the fault at the hypocenter of the 1966 main shock on the basis of patterns in the first motion data in the 1966 aftershocks. We find that this pattern also exists in the first-motion data from 1969 to 1987, but the 5° bend was not evident in the hypocentral distribution. This suggests that a more complicated explanation is needed to explain the first-motion data. Fault plane solutions were determined for the 71 events and 69 of these were compatible with strike-slip motion on a vertical San Andreas fault. An event located in the north end of the study area co-locates with the strike-slip solutions and may be a thrust or oblique solution. The other earthquake, located 2½ kilometers northeast of the fault, has a thrust or NNE-SSW striking right lateral solution but can not be explained by a San Andreas style mechanism. Both possible solutions can be explained by structures observed in the geology.