How Reliable Is Sitemeter?

I was interested, but more than a little confused, by the recent bloggy discussion of sitemeter stats and the burning question of whether Daily Kos's traffic is over-estimated, or not. My take over the years has been to see the data as somewhat dicey - rough at best. But I didn't realize quite how rough. I recently re-upped my subscription to Sitemeter and paid more money for more stats. Immediately, the balance between "visits" and "pageviews" shifted, with a pretty similar pageview count but a marked drop in visits as a proportion. More interestingly, the visit length increased dramatically as well - and now it tells me that the average visit length to the Dish is almost four minutes - far more than it once was and far more than most other blogs in its category. (To give a comparison, my average visit length is given as 3:58; Glenn's is 0.05; Powerline is 0.03; Kos is 0.22.) I can see why my average visit length would be higher than a quick-link site like Glenn's, but not that much longer. And I don't think a big shift with my new subscription was coincidental. The data from the Atlantic's measurements show no such change at all. So what gives?

My assumption has always been that these figures are pretty shaky. At best, I thought I could rely on them for some internally consistent numbers over time - a very rough gauge of progress. Now I'm not even sure about that. Any thoughts?