In the last three decades, the Ultra Steep Spectrum technique has been exploited by many groups since it was demonstrated that radio sources with very steep spectra (alpha < -1.0; S ∝ v(α)) are good tracers of high-z radio galaxies (HzRGs; z > 2). Though more than 150 HzRGs have been discovered up to now with this technique, little is known about its real effectiveness, as most of the ongoing searches still have incomplete follow-up programs. By selecting a new appropriate sample of USS sources from the MRC survey, the true searching efficiency of the USS technique has been quantitatively demonstrated for the first time in this paper. Moreover, it was compared with that of an optical search of HzRGs based on a simple cut of the galaxies r-band magnitude distribution. When no bias other than the radio-spectrum steepness is applied, the USS technique may be up to four times more efficient in selecting HzRGs with respect to an optical search. Nevertheless, when the search is limited to objects fainter than the POSS-II plates (r similar to 21), the USS technique is still 2.5 times more efficient (epsilon(USS) = 0.52 vs. epsilon(OPT) = 0.19). For an optical search to reach a comparable efficiency it is necessary to select objects fainter than r = 23, but this implies that about half of the HzRGs are lost because of the imposed magnitude bias. The advantage of the USS technique is that a similar to 0.5 search efficiency is already reached at the POSS-II plates limit, where all the optical identification work is done without telescopes. However, this technique has the drawback that up to 40% of the HzRGs of the sample are lost simply because of the applied spectral index bias. Interestingly, the introduction of a strong angular-size bias such as theta < 15&DPRIME; can double the searching efficiency irrespectively of the adopted technique, but only in the case that no optical bias has been introduced first. (C) 2003 Elsevier B.V. All rights reserved.