The smallest win your testcan actually prove.
Every A/B test has a floor: below a certain lift, your traffic simply can't tell signal from noise. This calculator finds that floor. Enter your baseline rate and visitors per variant and it tells you the minimum detectable effect — the smallest improvement you're allowed to believe.
Your numbers
Your current conversion rate for the page or flow you're testing.
How many people will see each version (control and variant) over the test.
The verdict
Min detectable lift (absolute)
0.9%
Percentage points the variant must move to be provable.
Min detectable lift (relative)
17.3%
The same floor as a % of your baseline.
Detectable variant rate
5.9%
Beat this rate to register a win.
Sample per variant
10,000
Visitors in each arm.
Where the bar sits
Standard is 95% confidence, 80% power. Tighten either and the MDE grows — you can prove only bigger lifts on the same traffic.
Sensitivity · traffic vs. resolution
What more visitors actually buy you
| Visitors / variant | Absolute MDE | Relative MDE | Detectable rate |
|---|---|---|---|
| 5,000 | +1.22 pts | 24.4% | 6.2% |
| 10,000 | +0.86 pts | 17.3% | 5.9% |
| 25,000 | +0.55 pts | 10.9% | 5.5% |
| 50,000 | +0.39 pts | 7.7% | 5.4% |
| 100,000 | +0.27 pts | 5.5% | 5.3% |
MDE scales with 1/√n: each row roughly quadruples traffic to halve the detectable lift. That's why chasing a tiny MDE with raw volume gets expensive fast — at a 5.0% baseline you'd need a huge sample to prove a sub-5.0% relative win.
Your move
Stop running tests that can't win.
At a 5.0% baseline with 10,000 visitors/variant (95% conf, 80% power), the smallest detectable lift is +0.86 pts (17.3% relative) — anything smaller is noise.
Send me your traffic and your testing roadmap. In 30 minutes I'll tell you which tests are actually powered to prove something, which ones are wishful thinking, and where to point the program so you stop burning cycles on inconclusive results. Free, and you keep the plan either way.
Plain English
What minimum detectable effect actually means
The minimum detectable effect (MDE) is the smallest difference between your control and variant that your test has a real chance of catching. It's not a number you choose for fun — it falls straight out of three things: your baseline conversion rate, your sample size per variant, and how strict you set your confidence and power. Below the MDE, a true lift can exist and your test will still call it a tie, because the result is buried under random noise.
Most teams get this backwards. They run a test, see "no significant difference," and conclude the variant doesn't work. Often the variant did work — just by less than the test could ever detect. If your MDE is 8% relative and your redesign delivered a genuine 4% lift, the test was statistically blind to it before a single visitor arrived. You didn't learn the variant lost. You learned your test was underpowered.
That's why you compute MDE before you launch, not after. It tells you the price of admission: at this traffic, only lifts of at least this size are worth testing for. If the smallest win worth shipping is below your MDE, you either need more traffic, a longer test window, or a bigger swing. This pairs directly with a sample-size calculator — MDE answers "what can I detect with the traffic I have," sample size answers "how much traffic do I need to detect the lift I want."
The formula
MDE (abs, pts) = (Zα + Zβ) × √(2 × p × (1 − p) / n) · p = baseline rate, n = visitors per variant
At a 5% baseline (p = 0.05) with 10,000 visitors per variant, 95% confidence (Zα = 1.96) and 80% power (Zβ = 0.8416): MDE = (1.96 + 0.8416) × √(2 × 0.05 × 0.95 / 10,000) × 100 ≈ 0.86 percentage points. That's a relative lift of ~17.2% (0.86 ÷ 5). So at this traffic, a variant has to push you to about 5.86% before the test can prove it. A move to 5.4% would be real and invisible.
Your MDE came back too high. Now what?
MDE is larger than any lift you realistically expect.
Your test is underpowered before it starts. Don't run it as-is — you'll almost certainly get an inconclusive result and waste the cycle. Get more traffic per variant or extend the window.
Baseline rate is very low (under ~1%).
Low baselines need enormous samples. The √(p(1−p)) term punishes rare events. Test higher up the funnel where the rate is bigger, or batch this test to run far longer.
Traffic is fixed and the MDE is still too big.
Stop testing tiny tweaks. Reserve this test for a bold swing — a full redesign, a new offer, a different page — that can plausibly clear the MDE. Small polish belongs on high-traffic pages only.
You need a smaller MDE fast.
Doubling traffic only shrinks MDE by about 30% (it scales with 1/√n). Loosening power from 90% to 80% helps a little. The cheapest real lever is raising the baseline rate you're testing against.
Eight ways to shrink your MDE (or live with it)
01Pool more traffic per variant
MDE falls with 1/√n. Four times the visitors halves the MDE. It's the most reliable lever, just not a cheap one.
02Run the test longer
If you can't get more daily traffic, extend the window. Two weeks at the same rate beats one week, as long as you cover full business cycles.
03Test higher in the funnel
A 30% click rate has a far smaller MDE than a 2% purchase rate at the same n. Move the test to where the baseline is bigger.
04Reduce variants
Every extra arm splits your traffic and inflates each variant's MDE. A clean A/B detects smaller lifts than an A/B/C/D.
05Accept 80% power
Don't reflexively crank power to 95%. The standard 80% gives a meaningfully smaller MDE and is the accepted norm for a reason.
06Swing bigger, not smaller
If the MDE is fixed at 10% relative, stop testing 2% tweaks. Ship changes bold enough to clear the floor or don't test them at all.
07Cut variance with targeting
Testing one clean segment instead of all traffic can lower variance and improve sensitivity — at the cost of generalizability.
08Pre-register the MDE
Write down the MDE before launch and refuse to ship sub-MDE "wins." It kills the temptation to chase noise after the fact.
The vocabulary
- Minimum detectable effect (MDE)
- The smallest true difference your test can reliably detect at your chosen confidence and power. Lifts below it tend to read as inconclusive.
- Absolute MDE
- The MDE in percentage points. From a 5% baseline, a 0.9pt absolute MDE means you can detect a move to 5.9%.
- Relative MDE
- The MDE as a percentage of the baseline. A 0.9pt move on a 5% baseline is an 18% relative lift. Most teams reason in relative terms.
- Confidence (1 − α)
- How sure you are a detected effect is real. 95% means a 5% false-positive rate. Higher confidence raises the MDE.
- Power (1 − β)
- The chance of catching a real effect of MDE size. 80% is standard. Higher power also raises the MDE on fixed traffic.
- Baseline rate (p)
- Your control's current conversion rate. It sets the variance √(p(1−p)) that drives how hard the effect is to detect.
MDE questions, straight answers
It's the smallest real lift your A/B test has a fair chance of proving, given your baseline rate, sample size, confidence, and power. Below the MDE, a true improvement will usually show up as "no significant difference," because the effect is smaller than the test's resolution. You compute it before launching to know whether the test is even capable of finding the win you care about.
Keep going
A calculator tells you what. A call tells you what to do about it.
Send me the account behind these numbers. I'll tell you straight where the money's leaking and what I'd fix first — free, and you keep it whether you hire me or not.