Context. The study of the oldest and most metal-poor stars in our Galaxy promotes our understanding of the Galactic chemical evolution and the beginning of Galaxy and star formation. However, they are notoriously difficult to find, with only five stars at [Fe/H] < −5.0 having been detected to date. Thus, the spectrophotometric data of 219 million sources which became available in the third Gaia Data Release comprise a very promising dataset for the identification of metal-poor stars. Aims. We want to use the low-resolution Gaia Blue Photometer / Red Photometer (BP/RP) spectra to identify metal-poor stars. Our primary aspiration is to help populate the poorly constrained tail of the metallicity distribution function of the stellar halo of the Galaxy. Methods. We developed a metal-poor candidate selection method based on flux ratios from the BP/RP Gaia spectra, using simulated synthetic spectra. Results. We found a relation between the relative iron abundance and the flux ratio of the Ca H & K region to that of the Hβ line. This relation is temperature and surface gravity dependent, and it holds for stars with 4800 K ≤ T eff ≤ 6300 K. We applied it to noisy simulated synthetic spectra and inferred [Fe/H] with an uncertainty of σ [Fe/H] 0.65 dex for −3 ≤ [Fe/H] ≤ 0.5 and G=15-17mag, which is sufficient to identify stars at [Fe/H] < −2.0 reliably. We predict that by selecting stars with inferred [Fe/H] ≤ −2.5 dex, we can retrieve 80% of the stars with [Fe/H] ≤ −3 and have a success rate of about 50%, that is one in two stars we select would have [Fe/H] ≤ −3. We do not take into account the effect of reddening, so our method should only be applied to stars which are located in regions of low extinction.