I don't dispute the theory here, but I'd sure love to see where the data came from. Via Dale Wolf, I see this:

According to Jim Lanzone, Senior Vice President of Search Properties at Ask Jeeves, there are 36,000 blogs that "really matter" since they have 20+ subscribers. That means only a small percentage of blogs have enough relevant material to gather a community around them.

The thing that is generally useful from Lanzone's analysis is that subscriptions for blog feeds through Bloglines gives us a relative understanding of the blog world:

Level 1 -- 36,000 blogs that are serious enough to garner 20 subscribers.
Level 2 -- 14,363 blogs that have +50 subscribers
Level 3 -- 437 blogs that have +1,000 subscribers
Level 4 -- 60 blogs that have +5,000 subscribers
Level 5 -- only one blog with +50,000 subscribers

Here's the gist of my question. I periodically go to BlogLines to see how many people are subscribing to my blog through BlogLines. Yesterday, that number was 302. Today, it tells me that there are zero with public profiles. So either BlogLines reset that data, everyone who was subscribed dropped off, or they all en-masse decided to hide their profiles. Occam's razor leads me to the first conclusion, but - the loss of that data tells you something about how slippery this area is.

So say my Bloglines subscribers are still around 300. When I looked at my logs today, I noticed that there are over 4600 unique IP addresses subscribed to one or more of the feeds on cincomsmalltalk (across all blogs). There are other online subscription services too - Newsgator comes to mind. I'm sure there are others I'm forgetting. The point is - I have no easy way of telling how big the community around the blogs on this server is. I can get a rough feel based on commenting frequency and log activity, but that's very rough - and the subscriber numbers are a lower bound anyway, due to proxy servers - if 10 people from Cincom subscribe to your feed, for instance, it will look like one unique IP.

Bottom line - does anyone have a methodology that captures anything that resembles real data on this stuff?