Response: This is based on the work of Erik Brynjolfsson, an MIT economist, and his team at the Sloan School. They fit a standard Pareto power law distribution to Amazon sales rank and summed the total above 40,000 (the amount of books stocked by the average bookstore). As you'll see in the paper above they did a really interesting job of forensic economics, buying multiple copies of books deep into the tail to see how it changed the sales rank and then using those figures to extract absolute sales figures from the rank numbers.
However, their research assumes two things: first, that the research they based their work on measured sales of the head (bestsellers) accurately; and second, that the Amazon sales rank algorithm is accurate. I'm told that in fact neither are quite right. Most analyses (such as this) tend to undercount sales at the very top of the list, and Amazon's calculations are still imperfect. This may account for some overcounting of the tail.
I've now spoken to Jeff Bezos (and others) about this. He doesn't have a hard figure for the percentage of sales of products not available offline, but reckons that it's closer to 25-30%. That would put it in line with Netflix's and Rhapsody's figures. This is still a large fraction of their business, and the fastest growing segment. I've updated the version of my article on ChangeThis accordingly.