In my experience working with content sites, the real bounce rate reduction you see in moving from shorter articles to longer ones is closer to 5% or so - from, say, 77% to 72%. Still dramatic, but not nearly as dramatic as the drop you'll get by double-inserting GA code.
From the article:
Stub stories work like this: You write the first installment like any other story. But when more news breaks, you go back to the article, insert an update at the top, and change the headline and subheadline to reflect the update. Our system updates the story "slug" when the headline changes--check the URL of this story, and you'll see words from the headline in the URL: /this-is-what-happens-when-publishers-invest-in-long-stories. But the number preceding the slug--on this article, it's 3009577--is a unique node ID which never changes. So essentially, every time we update an article, we get a fresh URL with a fresh headline, but pointing back to the same (newly updated) article.
We can check this by checking older stubs.
Edit: on checking the page source, I only see GA in there 2x - once to call a GA API from google-analytics.com, and once more to (I think) load ga.js from google-analytics.com. How many times do you see that page calling GA?
So if you're doing things like event tracking, without counting the event as a page view, then you'll still see another request to _utm.gif
They appear to have one normal _trackPageView call in the header, then one or more other calls coming from within minimized JavaScript files.
Here's one:
_gaq.push(["_trackEvent", e, t, n.toString(), parseInt(r, 10), !0])
_trackEvent(category, action, opt_label, opt_value, opt_noninteraction)
Presuming the other calls are identical, then their reported bounce rate should be correct.
However, the top comment references Google Analytics events on scroll that don't appear to be there now. One of FastCo.Labs folks also made the following comment:
This made me curious as well. It turns out there was a
technical change made at around the same time that is
going to account for some of it. We're running an
experiment now and will update with results.