Across the 1,140 winning A/B tests in LiftArchive that report a percentage lift, the median lift is +19% and the mean is +43.7%.
For single-change ecommerce A/B tests measured on revenue per visitor, the median win is a much smaller +10.0%.
These are lifts on published winners, so the typical test you run, including flat and losing ones, will land well below them.
This page shows how we counted, how lift varies by business type and metric, and why published lifts run high.
If you want the share of tests that win in the first place, see our A/B test win rate research.
Average uplift in an A/B test: what the database shows
How we counted
We started with all 1,724 results in the LiftArchive test database as of November 2026. We kept a result only if its headline outcome contains a signed percentage, such as "+7.8% revenue per visitor" or "-9.17% revenue per visitor".
That left 1,355 quantified results: 1,149 marked won, 136 flat and 70 lost.
We took the first signed percentage in each result as its lift.
Results stated only as "2x" or in dollars were left out.
Nine winners reported a falling number as the good outcome (lower cost per acquisition, for example), so we dropped them from the winner figures.
That leaves 1,140 winners with a positive lift.
Each result is one test, not a weighted sample.
A +956% click-through change counts the same as a +0.9% revenue change, which is why we lead with medians.
The headline numbers
| Group | n | Median lift | Mean lift | Middle 50% (25th-75th percentile) |
|---|---|---|---|---|
| All quantified winners | 1,140 | +19.0% | +43.7% | +8.4% to +45.4% |
| Winners from single A/B tests only | 855 | +14.5% | +33.4% | +7.0% to +33.1% |
| Ecommerce A/B tests, revenue per visitor | 200 | +10.0% | +17.3% | +6.3% to +19.4% |
| Flat (inconclusive) results | 136 | +5.2% | +14.8% | +1.7% to +12.5% |
| Losing results | 70 | -9.2% | -13.1% | -16.3% to -5.4% |
The mean is more than double the median because of a long right tail.
One in ten winners (121 of 1,140) claims more than +100%, and the 90th percentile is +105%.
The gap between the first two rows matters.
Results marked "bundled" (several changes shipped together, such as a full landing page redesign) have a median win of +35.4%, and before-and-after comparisons have a median of +27.0%.
Neither is a clean A/B test of one idea, so treat the single-test figures as the better guide.
How the winning lifts are spread
| Lift reported | Winners | Share |
|---|---|---|
| Up to +2% | 33 | 3% |
| +2% to +5% | 110 | 10% |
| +5% to +10% | 215 | 19% |
| +10% to +20% | 248 | 22% |
| +20% to +50% | 281 | 25% |
| +50% to +100% | 132 | 12% |
| Over +100% | 121 | 11% |
Only 13% of published winners claim less than +5%.
In a real testing program, small effects are common, and most tests that try to detect them come out flat.
Their absence here is the first sign of selection, covered below.
What a median win looks like
Wins near the ecommerce median are ordinary changes.
Dave Diederen's test of a new first product image on the product page returned +10.44% revenue per visitor.
Blend Commerce's exposed mobile search for Oak Tree Memorials returned +9.93% revenue per visitor.
At the small end, Life is Good's free shipping threshold change from $25 to $35 returned +1.3% revenue per visitor, and DRIP Agency's removal of optional checkout fields for KoRo returned +0.9% revenue per user.
On a large store, lifts that size can be worth a lot, but they need big samples to detect.
@DaveDiederen
Blend Commerce
DRIP AgencyTypical conversion lift by business type and metric
The typical lift depends heavily on what you sell and what you measure.
The table below uses winners from single A/B tests only, to keep bundled redesigns from inflating the figures.
| Vertical (single A/B tests) | Winners | Median lift | Mean lift | Middle 50% |
|---|---|---|---|---|
| Ecommerce | 667 | +11.9% | +22.9% | +6.2% to +25.4% |
| Lead generation | 85 | +31.0% | +68.9% | +14.9% to +66.0% |
| SaaS | 76 | +34.0% | +82.4% | +16.8% to +99.9% |
| Mobile apps | 27 | +25.0% | +43.4% | +10.0% to +50.0% |
Lead-gen and SaaS wins look about three times larger than ecommerce wins.
Part of that is real: a form or sign-up page has one action, and removing a step can move it a lot.
Kareo's shorter sign-up form returned +30% new doctor sign-ups, and Reassured's multi-step life insurance quote form returned +31.23% form submissions.
Part of it is the data. Our ecommerce winners are dominated by recent, single-change tests from agencies that publish many small results.
Many lead-gen and SaaS results are older public case studies, which are more likely to be the standout tests a company chose to share.
The metric changes the number
| Metric reported (all winners) | Winners | Median lift | Middle 50% |
|---|---|---|---|
| Revenue per visitor | 205 | +10.1% | +6.4% to +20.3% |
| Add-to-cart rate | 30 | +15.5% | +6.8% to +26.0% |
| Revenue or order value | 107 | +16.9% | +6.1% to +33.2% |
| Conversion rate | 387 | +20.0% | +9.0% to +44.5% |
| Leads, sign-ups or trials | 111 | +44.0% | +20.0% to +100.0% |
| Clicks or engagement | 67 | +53.1% | +23.0% to +159.0% |
We grouped metrics by keywords in each result, so the groups are approximate.
The pattern is still clear: the closer a metric sits to money, the smaller the lift.
Clicks are easy to move, and a big click-through gain often shrinks by the time it reaches orders.
The largest single figure in the database is VMware's CTA moved above the fold at +956% click-through.
MECLABS Institute's long versus short Medicare landing page reported +638% leads.
Results like these are real for their pages, but they say little about what your next test will do.
Our biggest A/B test wins page collects them.
If your goal is revenue, judge tests on revenue per visitor.
For ecommerce, the honest typical win on that metric is in the high single digits to low teens.
Why published lifts run high
Public test results skew toward wins. In our database, 84% of all 1,724 results are marked won (1,447).
Controlled studies of whole experimentation programs find far fewer winners.
At Microsoft, Ron Kohavi and colleagues reported that "only about 1/3 of ideas improve the metrics they were designed to improve," with the rest flat or negative (Kohavi et al., Online Experimentation at Microsoft).
Georgi Georgiev's analysis of 1,001 A/B tests run on Analytics-Toolkit.com found a median lift of 0.08% across all tests and a mean of 2.08%.
Among the 33.5% that were significant winners, the median lift was 7.5% and the mean 15.9%.
That 7.5% median is a fair benchmark for an unselected set of winners, and it sits below our +10.0% ecommerce revenue-per-visitor median and well below our +19% overall median. Three forces explain the gap:
- Selection. Teams and agencies publish the tests they are proud of. Clean Commit is an exception: it publishes its whole library, and of its 173 results here only 16 won, while 35 lost and 122 were flat.
- The winner's curse. A test that clears significance with a small sample usually does so because its observed lift came out high by chance. The true effect tends to be smaller, and it can shrink further after launch through regression to the mean and the novelty effect.
- Bundled changes and soft metrics. Redesigns and click metrics produce the biggest numbers, as the tables above show.
Flat results make the same point from the other side.
The median flat result in our data still shows a +5.2% point estimate.
Clean Commit's simplified product page showed +26.5% conversion but was not significant, which is why a raw lift without its confidence is not evidence of a win.
Losses are real too. Dave Diederen's removal of an email popup cost -20.51% revenue per visitor, and a countdown timer for a non-US European market cost -35.61%.
What lift to plan for
For planning, use a conservative figure. If you run single-change ecommerce tests and measure revenue per visitor, assume a typical winner is worth around +5% to +10%, and that most tests will not win at all.
That assumption should set your minimum detectable effect.
If your sample only lets you detect a +20% change, most real wins will look flat.
Check the traffic you need before you start.
Do not budget for the mean. A +43.7% average is pulled up by a small number of extreme results, many of them redesigns or click metrics.
The median, and the middle 50% range, describe what a normal win looks like far better.
When a vendor or agency quotes a big average lift, ask three things: how many tests it covers, whether flat and losing tests are included, and which metric it measures.
For more of our analysis, browse the research hub.