The sensitivity of the Streckeisen's very broadband (VBB) accelerometers is routinely measured on tilt tables by the manufacturer, with an announced accuracy of about 1 per cent. Nevertheless, the transportation of the station or different in situ environmental conditions may modify the sensitivity. As one may expect that a high accuracy in the wave amplitude will be required in the future by seismologists, we developed and tested an in situ absolute calibration method, which does not require the seismometers to be moved. Its principle is simple: A mass is moving in the vicinity of the force-balance accelerometer, and the perturbation of the gravitational field is measured. This calibration method, because it requires the use of G, the gravitational constant, is termed the G-calibration. At a distance of 0.5 m, the gravitational field of a 30 kg mass is 8 × 10−9 m.sec−2, which is 2 order of magnitude greater than the instrumental noise. At the test station of SBB, this acceleration is still 50 times above the seismic noise level for the vertical component, but little above the seismic horizontal noise. The calibration system consists into a small telescope platform (diameter 0.5 m) supporting a horizontal bar of 1 m. Two metallic cylinders of equal mass (about 30 kg) are placed at the two ends of the bar, symmetrically with respect to the vertical rotation axis of the platform. The rotation velocity of the system is controlled with a high accuracy. The accelerometers are located at less than 1 m from the system axis. For the vertical component, we measured the sensitivity at various periods between 2000 and 50 sec, and the result was within 10 per cent of the specified sensitivity; this difference was of the same order as the expected accuracy. For the horizontal component, a slight asymmetry of the two masses with respect to the rotation center and an irregular coupling of the platform to the ground induced a periodic variation of the tilt, dominating over the gravitational signal of the masses at the frequencies of interest, which made the sensitivity measurement less accurate. Nevertheless, the result was still close to the announced sensitivity. We expect that an improved version of our calibration system will allow an accuracy of 1 per cent for the vertical, with a few hours of calibration. This will require an accuracy of a few millimeters in the geometry of the calibration system. Improving the results for the horizontal component will require a careful installation in order to eliminate the tilt perturbation and should lead to an accuracy of a few per cent.