Greg Linden used to work at Amazon. From an older blog post of his:
"In A/B tests, we tried delaying the page in increments of 100 milliseconds and found that even very small delays would result in substantial and costly drops in revenue."
I find it immaterial considering the easier thing to speed up is the number of HTTP requests on the site, and that has clearly gotten worse over the years, not better. I pushed heavily to get HTML cleaned up and valid (showing an easy 10-20% reduction with HTML alone), reduce the amount of javascript, etc. It simply wasn't a priority. To be fair, at the time I was there the priority was to grow as fast as possible and deal with profits later.
The power point linked to in the article implies that its from Amazon's own internal data, but unfortunately there's very little substance presented. However, I don't find it surprising that amazon found a high correlation between responsive and overall sales.
The source slide actually says "+100 ms -1% sales". A 1% drop in sales would likely cost even more in profits, given fixed costs.
In either formulation, it's within the range of plausibility.
In my experience, a single change that makes a web server faster causes a slow, steady increase in traffic over the next few days/weeks, as the audience reacts to the new responsiveness.
Note also that Google considers page load latency in scoring ad relevance, and almost certainly natural search result ranking as well. So everything improves with speed.
I doubt the dropoff curve is linear. It's probably negligible up to a small threshold (500 ms?) with exponential decay in the tail. I'm willing to believe that a 1-second delay is a 10% hit, though I wonder how that would be measured, but this doesn't necessarily entail that 100ms costs 1%.
