Background: Transcription factor binding sites (TFBS) impart specificity to cellular transcriptional responses and have largely been defined by consensus motifs derived from a handful of validated sites. The low specificity of the computational predictions of TFBSs has been attributed to ubiquity of the motifs and the relaxed sequence requirements for binding. We posited that the inadequacy is due to limited input of empirically verified sites, and demonstrated a multiplatform approach to constructing a robust model. Results: Using the TFBS for the estrogen receptor (ER)alpha (estrogen response element [ERE]) as a model system, we extracted EREs from multiple molecular and genomic platforms whose binding to ERalpha has been experimentally confirmed or rejected. In silico analyses revealed significant sequence information flanking the standard binding consensus, discriminating ERE-like sequences that bind ERalpha from those that are nonbinders. We extended the ERE consensus by three bases, bearing a terminal G at the third position 3' and an initiator C at the third position 5', which were further validated using surface plasmon resonance spectroscopy. Our functional human ERE prediction algorithm (h-ERE) outperformed existing predictive algorithms and produced fewer than 5% false negatives upon experimental validation. Conclusion: Building upon a larger experimentally validated ERE set, the h-ERE algorithm is able to demarcate better the universe of ERE-like sequences that are potential ER binders. Only 14% of the predicted optimal binding sites were utilized under the experimental conditions employed, pointing to other selective criteria not related to EREs. Other factors, in addition to primary nucleotide sequence, will ultimately determine binding site selection.