Natural protein binding sites are often the most “druggable” sites on proteins, while alternative protein surfaces can be difficult targets. To explore the structural basis of this phenomenon, we used synthetic coevolution to engineer new interactions between naïve surfaces, simulating the de novo formation of protein complexes. We isolated seven distinct structural families of protein Z-domain complexes and found that synthetic complexes explore multiple shallow energy wells through ratchet-like docking modes, while complexes co-evolved from a natural binding surface converged in a deep energy well with a relatively fixed docking geometry. Epistasis analysis using machine learning to estimate fitness landscapes extracted “seed” contacts emerging from silent surfaces between binding partners that anchored the earliest stages of encounter complex formation. These data suggest why natural binding sites attract binders: alternative surfaces have a shallow energy landscape that disfavors tight binding, likely due to evolutionary counter-selection. Our findings have implications for understanding druggable versus undruggable surfaces.