Here we present a portable X-ray fluorescence (pXRF) dataset collected in situ (n = 1591) and a laboratory dataset (n = 226) from a soil sampling campaign in Marirongoe, Mozambique, to document the strength of rapid geochemical data collection in the field during mineral exploration. Real-time mapping of the geochemistry of underlying granite by utilising pXRF analysis of soil samples identified variation in granitic composition, thus allowing exploration to rapidly focus on the most prospective areas for Ta-Nb-U-REE mineralization. Principal components analysis and clustering protocols are applied to the centred log-ratio transform of a selection of eight elements (Ca, Fe, K, Rb, Sr, Ti, Mn, and Zr) to identify rocks of the same geochemical affinity a posteriori. Maps of these clusters reveal a map pattern that provides an interpretation of the underlying geology. One of the limitations of pXRF is false elemental concentrations being detected due to spectral overlaps between elements. We provide a possible solution to this problem through statistical data analysis using a probabilistic modelling approach. We propose a binary approach whereby the pXRF data for these elements, such as Sn, can be considered in the context of presence (detected; >150 ppm) or absence (not detected; <150 ppm) as comparison to laboratory data shows that the concentration of Sn is reliably detected at concentrations >150 ppm. A kernel density estimator and Bayes conditional probability can provide an effective method for calculating the probability of a sample having elevated content of elements, such as Sn, which may be variably detected by pXRF (depending on matrix and concentration). Utilising statistical approaches to treat large geochemical datasets, such as those that can be generated by pXRF, as they are collected, can provide timely and significant insights that might otherwise not have been apparent in elemental concentration maps alone.