One of the most crucial problems with genome-wide experimental analysis is how to extract meaningful biological phenomena from the resulting large data sets. Here, we present modeling and prediction techniques that are applied to genome-wide identification of in vivo protein–DNA binding sites from ChIP-based data sets. We develop a simple mixture probabilistic model of occurrence of non-specific and specific TF–DNA binding events for transcription factor binding to any site in the genome. We calculated the statistical significance of specific and non-specific random binding events using Kolmogorov–Waring and exponential functions, respectively. The binding events in the chromosome regions associated with non-specific, non-random binding loci were also identified and filtered out. The mixture model fits equally well to five different TFs (ERE, CREB, STAT1, Nanog, Oct4) data provided by ChIP-PET, SACO, and ChIP-Seq methods included in this study. We present a uniform methodology for estimating specificity, total number of binding sites, and sensitivity of data sets detected by these ChIP-based genome-wide experimental systems. We demonstrate strong heterogeneity of specific TF–DNA binding sites in terms of their avidity and by correlation between observed relative binding avidity of specific TF–DNA binding site and the level of mRNA transcription of the nearest gene target. Finally, we conclude that the sensitivity problem has not been resolved by current ChIP-based methods, including ChIP-Seq.