Quantum systems can be used to certify the existence of true randomness in nature, and semi-device-independent (SDI) randomness expansion has been used to achieve this goal under the prepare-measure configuration. Usually, randomness is certified with the degree of violation of some inequalities. However, the observed statistics characterizing SDI protocols contain much richer information than just a single value of the inequality. Thus, there might be a gap between the randomness in observed statistics and the randomness we certify. The purpose of this paper is to optimize the use of the full observed statistics to reduce that gap. It shows that certifying randomness directly from the full observed statistics but not some inequality is available in the SDI protocols. Besides, average of guessing probabilities should be considered as randomness quantitative standards, since it certifies more randomness than the previous maximal guessing probability method. In the analysis, we take into account the assumption that the devices and the potential adversary are classically correlated.