I think you also should consider the very small sample data set your using. In the mainstream stuff I was doing, we'd send hundreds of thousands of hits before even determining a ratio. Less sample data = greater fluctuations = less real understanding.
