The Amazon 100ms Latency Study: Where the “1% of Sales” Figure Comes From
“Every 100ms of latency cost Amazon 1% of sales” is the most-cited statistic in web performance. It is real, but it is also routinely misattributed and misquoted. Here is exactly where it comes from, what the experiment actually measured, and the independent studies that back it up.
The short answer
The statistic traces to Greg Linden, a former Amazon engineer, who shared it on his blog Geeking with Greg and in his 2006 Stanford guest lecture “Make Data Useful” (the “Speed Matters” slide). Amazon ran A/B tests that deliberately delayed pages in 100-millisecond increments and measured the revenue impact. Linden reported that every 100ms of added latency cost roughly 1% of sales — sales, not profit. Amazon never published the experiment formally, so it is best read as a credible first-hand account rather than a peer-reviewed study.
What the experiment actually measured
Amazon did not measure a slow site against a fast one in the wild. Instead, engineers took the existing fast experience and artificially injected delay for a random subset of users, in steps of 100 milliseconds, then compared revenue per visitor between the delayed and control groups. This is a clean controlled experiment: the only difference between groups was the added latency, so the revenue difference can be attributed to speed alone.
The result, as Linden recounted, was that even these small, deliberately introduced delays produced “substantial and costly” drops in revenue, on the order of 1% of sales for every 100ms. The effect was large enough that Amazon treats latency as a first-order business metric, not just an engineering one.
Three things people get wrong about it
“1% of profit”
Linden said 1% of sales (revenue). The profit phrasing is a later misquote and implies a much smaller dollar figure.
“A published Amazon study”
Amazon never formally published it. The source is an engineer's talk and blog, which is why citations should name Greg Linden, not 'an Amazon study'.
“Still exactly 1% for everyone”
The 1% is specific to Amazon's 2006 site and audience. The relationship generalizes; the precise number does not.
The studies that corroborate it
The Amazon figure does not stand alone. Independent experiments at other companies found the same direction and a similar order of magnitude.
| Company | Year | Finding | Source |
|---|---|---|---|
| Akamai / SOASTA | 2017 | A 100ms delay can reduce conversions by up to 7%; a 1-second delay by up to 22%. | State of Online Retail Performance Report (Spring 2017, published 18 April 2017), analyzing roughly 10 billion user visits. |
| Bing (Microsoft) | 2009 | A 2-second slowdown reduced revenue per user by 4.3% and queries by 1.8%. | Eric Schurman, Microsoft, presented at the Velocity conference (joint Bing/Google session). |
| 2006 | Adding 0.5 seconds to search results reduced traffic by 20%. | Marissa Mayer, Web 2.0 Summit presentation. | |
| Deloitte + Google | 2020 | A 0.1-second mobile speed improvement lifted retail conversions by 8.4%. | Deloitte Digital, 'Milliseconds Make Millions' report. |
Frequently Asked Questions
What is the source of the Amazon '100ms latency = 1% of sales' statistic?
The figure comes from Greg Linden, a former Amazon engineer, who described it on his blog 'Geeking with Greg' and in his 2006 Stanford guest lecture 'Make Data Useful' (the 'Speed Matters' slide). Amazon ran A/B tests that deliberately delayed page rendering in 100-millisecond increments for a subset of users and measured the revenue impact. Linden reported that every 100ms of added latency cost roughly 1% of sales. Amazon never published the experiment formally, so the figure is best treated as an engineer's first-hand account rather than a peer-reviewed result.
Did the Amazon study say 1% of sales or 1% of profit?
Sales (revenue), not profit. Greg Linden's original statement was that every 100ms of latency cost about 1% of sales. The widely circulated 'cost 1% of profit' phrasing is a misquote that appeared in later retellings. The distinction matters: at Amazon's scale a 1% revenue effect is far larger in absolute dollars than 1% of net profit would be.
Is the Amazon '100ms = 1% sales' study from 2006 still accurate today?
The exact 1% figure is specific to Amazon's site, audience, and 2006 page architecture, so it should not be applied verbatim to other businesses. But the underlying relationship, that small latency increases produce measurable revenue loss, has been reproduced repeatedly since: Akamai/SOASTA (2017) found a 100ms delay could cut conversions by up to 7%, Google/Bing experiments in 2009 found a 2-second slowdown reduced Bing revenue per user by 4.3%, and Deloitte's 'Milliseconds Make Millions' (2020) found a 0.1s mobile speed improvement lifted retail conversions by 8.4%. The direction and rough magnitude have held for 20 years.
How much would 100ms of latency cost Amazon in dollars?
It scales with revenue. In 2006, 1% of Amazon's roughly $10.7B annual net sales was about $107M, so the original experiment implied that figure per 100ms. Against Amazon's recent annual revenue of more than $600B, 1% is over $6B. These are illustrative extrapolations of the original per-100ms relationship, not figures Amazon has published.