Experimental techniques to determine paleomagnetic field intensity are based on a theoretical framework that is valid only for single-domain (SD) ferromagnetic particles. Yet, most of the available materials exhibit distinctly non-SD properties. Designing the optimal paleointensity methodology for non-SD is, therefore, a fundamental challenge in paleomagnetism. The objective of this study is to experimentally test the IZZI Thellier absolute paleointensity method on small MD recorders. The test has two purposes: 1) to describe the characteristic non-SD patterns occurring in Arai plots, and 2) to identify the optimal approach in interpreting non-SD behavior. We carried out paleointensity experiments on 40 specimens from 4 synthetic re-melted slag samples with identical magnetic properties (mineralogy, texture, and non-SD state) produced under different field intensities. We ran three batches of IZZI experiments using different conditions that allow for a detailed characterization of the non-SD behavior. We find that the curvature of the Arai plot is systematically dependent on the angle and the proportion between the field used in the paleointensity experiment (B(TRM)) and the field in which the NRM was acquired (B(NRM)). Straight-line Arai plot occur when the two fields are parallel and equal, and seems to always give the 'true' slope. Convex curves occur when B(TRM) is parallel and significantly stronger than B(NRM). Concave curves occur in all the other cases and yield two end-case slopes that are always different than the 'true' slope. In addition, zigzagged patterns increase with the angle the proportion between B(TRM) and B(NRM). We test the accuracy of the 'best fitting' line approach and conclude that 'best fitting' line in curved plots cannot provide robust paleointensity estimates. Yet, the two 'end-case' slopes in concave curves provide adequate constraints for the true value. We introduce a new procedure to calculate a 95% confidence interval of the paleointensity from curved plots using bootstrap statistics. We substantiate the new procedure by conducting two independent tests. The first uses synthetic re-melted slag produced under known field intensities - 3 SD samples and 4 non-SD samples. The second compares paleointensity determinations from archeological slag samples of the same age - 34 SD samples and 10 non-SD samples. The two tests demonstrate that the bootstrap technique may be the optimal approach for non-ideal dataset. (C) 2011 Elsevier BM. All rights reserved.