Astrometry at or below the microarcsec level with an imaging telescope assumes that the uncertainty on the location of an unresolved source can be an arbitrarily small fraction of the detector pixel, given a sufficient photon budget. This paper investigates the geometric limiting precision, in terms of CCD pixel fraction, achieved by a large set of star field images, selected among the publicly available science data of the Transiting Exoplanet Survey Satellite (TESS) mission. The statistics of the distance between selected bright stars (G ≃ 5 mag), in pixel units, is evaluated, using the position estimate provided in the TESS light curve files. The dispersion of coordinate differences appears to be affected by long term variation and noisy periods, at the level of 0.01 pixel. The residuals with respect to low-pass filtered data (tracing the secular evolution), which are interpreted as the experimental astrometric noise, reach the level of a few milli-pixel or below, down to 1/5900 pixel. Saturated images are present, evidencing that the astrometric precision is mostly preserved across the CCD columns, whereas it features a graceful degradation in the along column direction. The cumulative performance of the image set is a few micropixel across columns, or a few 10 μpx along columns. The idea of astrometric precision down to a small fraction of a CCD pixel, given sufficient signal to noise ratio, is confirmed by real data from an in-flight science instrument to the 10−6 pixel level. Implications for future high precision astrometry missions are briefly discussed.