The U.S. Geological Survey has been using the Global Positioning System (GPS) to perform repeated surveys of networks of different sizes since January 1986. The intersite distances from these various networks are 223 km for a single line connecting the Palos Verdes and Vandenberg sites in southern California; 30–50 km for a network of three sites south of the San Francisco Bay near Loma Prieta; 9–30 km for a network located near Hebgen Lake, Montana; 7–12 km for a network of sites in the vicinity of Parkfield, California; and 240 m for a line located at the Goldstone antenna complex in the Mojave desert. The lengths quoted do not include the intersite distances for the “fiducial” sites included in the solutions for orbit determination, which can be up to 8000 km. We have attempted to assess the “long-term” precision and accuracy of determinations of relative site position for a range of line lengths. This assessment has consisted of several different studies. In assessing precision, we have generally investigated repeatability, usually about a best fit linear temporal trend. The smallest fractional repeatability was yielded by the measurements on the Palos Verdes to Vandenberg line. The root-mean-square (rms) residual about the best fit line for these determinations was 6 mm (0.03 parts per million, or ppm) for the north component, 11 mm (0.05 ppm) for the east component, and 40 mm (0.18 ppm) for the vertical component. For these determinations, GPS data from 3–4 consecutive days were combined to produce a single estimate of the vector separation. We also used the day-to-day scatter from these measurements to perform an analysis of variance and found that these multiday determinations can be significantly more precise than estimates determined from a single observing session. The worst results were obtained for the Loma Prieta network, where we obtained rms residuals of 11–12 mm (0.3–0.4 ppm) for north, 14–18 mm (0.3–0.6 ppm) for east, and 18–21 mm (0.4–0.7 ppm) for vertical. We cannot definitely explain these poor results, but we have evidence to indicate that this poor scatter is due to the short observing session (∼4.5 hours, compared to ∼6.5 hours for the Palos Verdes to Vandenberg results). In assessing accuracy, we compared the GPS results to those from independent techniques. At Parkfield we compared the GPS estimates of fault-parallel motion to those from creep meters and alignment arrays and found excellent agreement (within 1–2 mm yr−1) not only for the slip but also for the variation of slip along the San Andreas fault. We compared the Loma Prieta and Hebgen Lake line length estimates to those obtained from a Geodolite, a high-precision laser distance-measurement device that the U.S. Geological Survey has used for 20 years for studies of crustal deformation. If we model the Geodolite minus GPS line length differences ΔL as ΔL = α + βL, where L is the line length, then we find α = 0.6 ± 0.5 mm and β = 0.1 ± 0.1 mm km−1. The uncertainties quoted are one standard deviation. From this comparison, we conclude that at this level of significance there is no detectable difference between lengths determined from Geodolite and GPS data. Finally, we compared the GPS estimates of the Palos Verdes to Vandenberg line length to those obtained from very long baseline interferometry (VLBI). The mean difference between the six GPS estimates of line length and a best fit line through the VLBI estimates was −1 mm, and the rms difference was 12 mm (0.05 ppm), close to the standard deviation of the GPS measurements for this predominantly east-west line.