This is exactly why I can't trust prescriptive conclusions from correlation studies. Didn't Google publish some method on resolving causation in correlation? Something about transfer of noise?