Although nutrient depletion and yield decline are common on sandy landscapes in upland Cambodia, small-scale farmers seldom use fertilisers partly due to lack of evidence on yield and economic benefits of fertilisers. Large-scale replicated experiments with multiple fertiliser types, rates, locations and seasons are costly, complex, high maintenance and are impractical in development settings. We developed a simple unreplicated on-farm experimental (OFE) method, each consisting of three short strips: T1 no fertiliser control, T2 15 kg P + 20 kg K / ha and T3 using twice the T2 rate. Peanuts was the test crop. The OFE were repeated among farmers, seasons and years. We completed 53 OFE on a sandy landscape in the early and main wet seasons of 2019-22. Cate-Nelson analysis of plots of relative yields against soil test values was used to determine critical values (soil test values below which a response is expected) for P and K. The yield data were skewed and analysed using non-parametric methods including paired t-test; Friedman’s rank test to determine whether at least one of the treatments significantly differed from the others; and Wilcoxon post hoc test to identify which specific treatment pairs differed. Yields of T2 were generally but not consistently better than T1. A linear least square regression model estimated a baseline yield (T1) of 1.51 and the effects of T2 and T3 of 0.62 and 0.68 t/ha respectively. These estimates were significant, but the model accounted for 18% of yield variations. The evidence points to significant but inconsistent yield benefits of T2. Bayes theorem calculated the posterior probability of profits given expert opinion, the probability of profits determined by the OFE and the probability of encountering deficient sites. A posterior probability of 85% of profits from adopting T2 is being communicated to the district to promote fertiliser use.