Minimum Detectable Effect Calculator
Already know how much traffic you have? This tells you the opposite of a sample-size tool: given the sample you can get, the smallest lift that test can actually prove — so you know whether the experiment is worth running at all.
Minimum detectable effect (MDE) is the smallest relative lift a test can reliably distinguish from noise with the sample you have. It is the inverse of a sample-size calculation: instead of asking ‘how much traffic do I need for this lift?’, it asks ‘given my traffic, how small a lift can I see?’ If your true improvement is smaller than the MDE, the test will likely report no significant difference even though the variant genuinely helps — so knowing the floor tells you whether a test is worth running.
Minimum Detectable Effect Calculator inputs and result
How to use this calculator
- Enter your baseline rateUse the control’s current rate. A lower baseline pushes the detectable floor higher for the same amount of traffic.
- Enter the sample you can get per variationBe realistic about the visitors each cell will collect in your window. This is the constraint that sets the floor.
- Set confidence and powerKeep 95% and 80% unless you have a reason to change. A stricter bar means only bigger lifts will register.
- Read the smallest detectable liftThe big number is the smallest relative lift the test can prove. Below it, real improvements will tend to read as ‘no difference’.
- Decide if the test is worth itIf your expected lift is below the floor, the test cannot settle the question — redesign for a bigger change or gather more traffic. Then export.
RGM Expert Says
MDE is the question most teams skip and then regret. They run a test, get a flat result, and conclude the idea failed — when in fact the test was never sensitive enough to see the lift they actually got. We run this calculation first so a null result is informative: if the floor is 15% and the change was always going to be a 4% improvement, the test was doomed before launch, and that is worth knowing on day zero.
It reframes the conversation from ‘will it win?’ to ‘can we even see it win?’ On a thin-traffic page, the honest floor might be a 25% lift — which means subtle copy tweaks are simply not testable there, and the team should either batch them onto high-traffic surfaces or commit to bolder swings. That single number saves weeks of inconclusive tests.
The inverse-square relationship is the part clients feel most. When they ask why they cannot detect a 5% lift on the traffic that comfortably detects 10%, the answer is that the second case needs about four times the sample. We use the floor to set expectations honestly: experimentation has a resolution limit, and it is set by traffic, not effort.
How it works
The detectable floor is found by inverting the two-proportion sample-size formula: we search for the smallest relative lift whose required sample equals the sample you actually have.
- N — the sample size per variation you can collect.
- p₁ — baseline rate of the control.
- zα, zβ — critical values for confidence and power.
- MDE — smallest relative lift whose required sample fits within N.
The floor inverts the standard two-proportion sample-size formula via a 64-step binary search on the relative effect — the same n(·) function used in our sample-size and budget calculators. Background on MDE and power: Evan Miller; Georgiev, Analytics-Toolkit.
Knowing the floor stops doomed tests before they start
Every test has a resolution limit. Below the minimum detectable effect, a real improvement is invisible — the noise is simply larger than the signal at that sample. Teams that do not compute the floor routinely run tests that could never have succeeded, then misread the flat result as proof the idea was bad.
The floor turns ‘run it and see’ into a feasibility decision. If a page can only detect a 20% lift and the realistic upside is 5%, the right move is not patience — it is a bigger swing, a higher-traffic surface, or a longer window. MDE makes that call before any budget is spent.
It also keeps null results honest. A non-significant outcome from a test with a sensible floor is real evidence the change did little. The same outcome from an underpowered test is evidence of nothing. The floor is what tells the two apart.
How sample size sets the floor
For a fixed baseline, the smallest lift you can detect falls as sample grows — but slowly, because of the inverse-square relationship. These figures use a 5% baseline at 95% confidence, 80% power.
| Sample per variation | Detectable relative lift | Read |
|---|---|---|
| 2,000 | ~41% | Only bold swings register |
| 10,000 | ~18% | Moderate changes detectable |
| 50,000 | ~8% | Subtle copy lifts in reach |
| 200,000 | ~4% | Near the practical resolution limit |
What the experts say about sensitivity
If you cannot detect the effect you expect with the traffic you have, you do not have an experiment — you have a hope. Compute the minimum detectable effect first.
Most flat A/B results are not failed ideas; they are tests that were never powered to see the lift that actually occurred.