Ah fair enough, I remembered seeing this sometime ago.

I think the disparity is pretty big as compared to the other statistics that you find around and the accuracy can be quite skewed, even today (4k vs 6.4k). That is taking into account that you're only looking at the listening nodes. Anyhow, he has never shared the methodology and the last time I checked, he preferred to keep it a secret so I can't really vouch or check for the accuracy.