Then I glanced at the referer chart and noticed that the source was stumbleupon.com. At once it all made sense.
Let me explain. The average StumbleUpon user turns to the service when they're bored, so bored they can't even go to the trouble of endlessly clicking on links on web pages like most of us do. Instead they click repeatedly on the "Stumble" button which takes them to random web pages supposedly somewhat tailored to their interests. They're not in it to read the articles (any more than someone who's flipping through the pages of Playboy is in it to read the articles). Instead they're in it for the tiny dopamine spike that they get each time they land on a new page.
Nine times out of ten, such a user will bounce immediately after looking at the title of your article, deciding that it's not something they're interested in. If they do start reading, a further nine times out of ten they'll bounce somewhere into the second paragraph. If you don't believe me, try using the product, and see how quickly you find yourself doing the same thing.
Before I go on to make my point, I should say that this is nothing more than a minor annoyance to me personally. I'm an academic; I'm not trying to monetize my site. And 33bits is a wordpress.com blog, so I don't pay hosting costs. The only reason I'm annoyed is that when I look at my stats page to see what sorts of articles my readers are most interested in, I have to mentally discount the articles that got StumbleUpon traffic. But anyone who pays hosting costs for their blog and is trying to make money (or spread an idea, or whatever) might want to take note of the following.
The architecture of StumbleUpon is fundamentally exploitative of the quid-pro-quo nature of free websites. A pageview from a StumbleUpon visitor costs just as much in bandwidth, but is a couple of orders of magnitude less likely to result in any sort of engagement. Your website wasn't meant to be viewed in a frame, so don't let it.Even though StumbleUpon has only 10 million users, this is a bigger problem than might seem at first sight. The recommendations that the system makes are voting-based, so the mechanics of popularity and the resulting traffic patterns are essentially the same as with Digg and Reddit, although the engagement numbers are very, very different. This means that most days you'll probably see no StumbleUpon traffic, but one day you'll get unlucky and the resulting spike will dominate your traffic and costs for that month, but you'll have nothing to show for it.
I would recommend some framebusting and User-Agent sniffing code to politely tell StumbleUpon users to go somewhere else, but whatever you do, don't put a Stumble button on your pages!
 I'm sure there's a sizeable fraction of users for whom the collaborative filtering aspect works well, and who consequently actually read the articles and engage with the sites. But even if half the users fall into this group (although I doubt it's anywhere near that high), most of the traffic generated by StumbleUpon users to any given site is going to be low quality because the dopamine junkies make 100x more clicks.