How often do A/B tests actually win, by how much, and how long do they take? We re-analyzed every qualifying test run on Stellar between 28 Feb 2025 and 13 Nov 2026, from 125 accounts. No cherry-picked case studies, just ranges.
Of the 445 qualified tests with a flagged control (28 Feb 2025 – 13 Nov 2026, our recalculation at 95% probability).
In our data, a variant beat the control with at least 95% probability in 16.9% of tests. The control won clearly in another 9.9%. The remaining 73.3% had no clear result, even after meeting our minimums of 300 visitors per version and 7 days.
That is in line with what larger testing programs publish. Kohavi and Chen (KDD 2024) report success rates of about 10–20% as the most common across companies that share them. So if roughly one test in six produces a clear winner for you, you are not doing it wrong. A clear loss is useful too: it stops you from shipping a change that would have hurt.
Data through 13 Nov 2026 · window 28 Feb 2025 – 13 Nov 2026 · n = 445 tests with a flagged control, 125 accounts in the qualified set.
Most lifts are small, and many are negative. Best variant vs control, relative change in conversion rate, across 445 qualified tests:
| Percentile | Lift vs control |
|---|---|
| 10th | −27.3% |
| 25th | −8.1% |
| Median | +1.0% |
| 75th | +13.0% |
| 90th | +53.1% |
Half of all tests landed between −8.1% and +13.0%. Treat any single "+50%" story with caution: in our data that happens in roughly the top tenth of tests, and very large swings mostly appear on pages with low baseline conversion rates (see below).
Data through 13 Nov 2026 · window 28 Feb 2025 – 13 Nov 2026 · n = 445.
Industry is self-reported by the account owner at signup. We only publish an industry when it has at least 10 qualified tests from at least 5 accounts, and no single account makes up more than half of the tests.
| Industry | Tests (accounts) | Median runtime (p25–p75) | Median visitors per version | Variant beat control | Lift vs control, p25 to p75 |
|---|---|---|---|---|---|
| Software | 109 (37) | 57 days (26–117) | 1,339 | 10.1% | −6.8% to +12.8% |
| Ecommerce | 97 (22) | 48 days (24–84) | 3,350 | 12.0% | −6.5% to +8.3% |
| Agency | 89 (25) | 56 days (31–109) | 1,995 | 23.6% | −6.6% to +23.9% |
| Other (self-reported) | 106 (26) | 85.5 days (38–184) | 1,736 | 20.8% | −11.2% to +5.8% |
| All qualified tests | 450 (125) | 57 days (29–116) | 2,016 | 16.9% | −8.1% to +13.0% |
"Variant beat control" is our recalculation at 95% probability, among tests with a flagged control. Agency accounts usually run tests on client sites, so that row mixes many kinds of businesses.
Data through 13 Nov 2026 · window 28 Feb 2025 – 13 Nov 2026.
Finance, education, nonprofit and government don't meet our thresholds yet (too few tests, too few accounts, or one account making up more than half). We'll add them when they do.
| Test type | Tests (accounts) | Median runtime | Median visitors per version | Variant beat control | Lift vs control, p25 to p75 |
|---|---|---|---|---|---|
| On-page edits (same URL) | 315 (94) | 62 days | 2,551 | 14.6% | −7.2% to +9.5% |
| Split URL (separate pages) | 135 (45) | 48 days | 1,397 | 22.1% | −10.9% to +22.2% |
Split URL tests compare two whole pages, so they tend to change more at once. In our data they also show a wider spread of outcomes, in both directions.
Data through 13 Nov 2026 · window 28 Feb 2025 – 13 Nov 2026.
| Versions in the test | Tests (accounts) | Median visitors per version | Some version reached 95% probability of being best | Variant beat control | Lift vs control, p25 to p75 |
|---|---|---|---|---|---|
| 2 (A/B) | 392 (115) | 1,838 | 27.8% | 17.1% | −9.0% to +10.1% |
| 3 or more (A/B/n) | 58 (30) | 2,666 | 17.2% | 15.5% | −1.0% to +23.5% |
Every extra version splits your traffic further. In our data, tests with three or more versions reached a clear result less often.
Data through 13 Nov 2026 · window 28 Feb 2025 – 13 Nov 2026.
Yes, a lot. Conversion here means the goal each test tracked, which can be a click, a signup or a purchase.
| Control conversion rate | Tests (accounts) | Lift vs control, p10 to p90 | Lift, p25 to p75 |
|---|---|---|---|
| Under 2% | 123 (56) | −53.1% to +167.1% | −19.2% to +53.8% |
| 2–5% | 76 (42) | −28.9% to +46.6% | −11.6% to +21.1% |
| 5–15% | 119 (46) | −25.3% to +30.7% | −6.8% to +9.8% |
| 15% and above | 127 (38) | −15.1% to +9.7% | −2.6% to +4.6% |
When a page converts under 2%, a handful of extra conversions shows up as a huge percentage. Big relative lifts on low-converting pages are common, and so are big relative losses. Plan with absolute numbers, not percentages alone.
Data through 13 Nov 2026 · window 28 Feb 2025 – 13 Nov 2026 · n = 445 tests with a flagged control.
Want to know what your own traffic allows? Plan your test with the sample size and duration calculator, or read how long to run an A/B test.
Data through 13 Nov 2026 · window 28 Feb 2025 – 13 Nov 2026.
Stellar reports each version's probability of beating the control, using Bayesian statistics. You see how likely a change is to be better, not a p-value.
Stellar Autopilot only calls a winner at 95% probability, after at least 7 days and 300+ visitors per version, the same bar we used for these benchmarks. See Autopilot
Edit a page with the no-code editor, or send traffic to two separate URLs. Both test types appear in these benchmarks.
See where visitors click and scroll on each version, so a flat result still teaches you something. Pair with Hotjar or Clarity
Pricing: Sandbox is $0 for up to 5k monthly tracked users. Plus is $29/mo billed annually for up to 25k, and Pro is $99/mo billed annually for up to 50k.
What is a good A/B test win rate?
In 445 qualified tests on Stellar (28 Feb 2025 – 13 Nov 2026), a variant beat the control with at least 95% probability 16.9% of the time, by our recalculation. Larger programs commonly publish 10–20%. A rate in that range is normal.
What percentage of A/B tests fail?
"Fail" usually means no clear winner. In our data, 73.3% of qualified tests had no clear result and 9.9% were clear losses for the variant. A clear loss still counts as a useful result.
What is a typical A/B test lift?
The median best variant was +1.0% vs control. Half of tests fell between −8.1% and +13.0% (n = 445, data through 13 Nov 2026).
Are these Stellar's declared winners?
No. We recalculated every test the same way: Bayesian, a uniform Beta(1,1) prior, unique visitors who converted, and a 95% probability of being best. This keeps tests comparable, whatever settings each team used.
Why don't you show results for my industry?
We publish a group only when it has at least 10 tests from at least 5 accounts, and no single account makes up more than half of it. Smaller groups could identify a customer, and their numbers would mostly be noise.
How often is this updated?
We plan to refresh the numbers monthly. Every figure on this page carries its data-through date and window.