Low traffic A/B testing: can your site test at all?

This is the diagnostic calculator, not the one for planning a single test. If you already know which effect you are chasing and want its runtime, use the A/B test duration calculator. If you want the smallest detectable effect from a sample that already exists, use the MDE calculator. If the question is how many tests fit in your year, use the test velocity calculator. The question here comes before all of them, and it is the one that stalls small teams: with the traffic I have, can I test anything?

How to use it

1. Enter the visitors per month who go through the flow you want to test (not your whole site: only the people who see the tested page).
2. Enter your current conversion rate on that flow. It moves the result more than traffic does.
3. Say how many weeks you are willing to run one test. Six weeks is an honest cap for most teams; twelve is the most that usually survives seasonality and site changes.
4. Adjust the number of variations (2 = A/B) and the relative lift you would like to detect.
5. Pick your rigor. The default (95% confidence, 80% power) is the convention; relaxing to 90% and 70% cuts the sample and raises the risk of error, which can be worth it in exploratory work.
6. Read the verdict and the scenario table: for lifts from +10% to +50%, it shows the sample each one demands and whether it fits your window.

How it works: the math behind it

The tool chains two calculations, both running in your browser. The first turns your window into available sample; the second turns that sample into the smallest effect it can separate from noise.

Where p is your current rate, n is the sample per variation, z_α is the critical value for confidence (1.96 at 95%, two sided) and z_β the one for power (0.84 at 80%). The way back, how much sample a specific lift demands, uses the same normal approximation for two proportions as our sample size calculator, so the numbers on both pages agree.

Worked example (it reproduces the default output)

With the values already filled in: 8,000 visitors/month, a 2% conversion rate, a 6 week window, 2 variations, a +20% target and standard rigor.

• Daily traffic: 8,000 / 30 = 266.7 visitors per day.
• Window: 6 × 7 = 42 days, so 42 × 266.7 = 11,200 visitors in total.
• Sample per variation: 11,200 / 2 = 5,600.
• Smallest detectable lift: (1.96 + 0.84) × √(2 × 0.02 × 0.98 / 5,600) = 2.80 × 0.00265 = 0.0074, that is 0.74 percentage points.
• In relative terms: 0.0074 / 0.02 = +37.1%. The variation would have to move from 2% to 2.74% for the test to call it.
• Verdict: tight. Only a radical change produces a lift that size.
• The +20% target would demand 21,109 visitors per variation, 42,218 in total, or 159 days (22.7 weeks). It does not fit in 6 weeks: you would need 30,156 visitors per month, about 3.8 times your current traffic.

The scenario table closes the argument: with these numbers only the +50% scenario fits the window (3,826 per variation, 4.1 weeks). The +30% row asks for 10.6 weeks, +20% asks for 22.7 and +10% asks for 86.6 weeks, more than a year and a half. That is exactly what the tool prints when the page loads.

How to read the verdict

The number that matters is the smallest detectable relative lift, because it tells you what kind of change is worth testing on your site:

• Up to +10%: comfortable. You are in the range where ordinary CRO tests produce results, so refinement hypotheses are fair game.
• +10% to +25%: workable, but only for structural changes. Test the offer, the value proposition, the whole page layout, not the button color.
• +25% to +50%: tight. Lifts that big exist, but they are rare. Consider swapping the metric for a more frequent micro conversion before burning six weeks.
• Above +50%: unfeasible on the final conversion. Testing like this buys an inconclusive result with weeks of your calendar. Change the approach.

One important warning: never use an underpowered test to conclude that a variation does not work. Low power means the test fails to detect real effects, so a non significant result is not evidence of no effect. To see the power of a test you already ran, use the statistical power calculator.

The plan B for low traffic sites

When the verdict comes back tight or unfeasible, there are honest ways out. All of them trade precision for learning speed, and none of them involve stopping the test when it looks pretty.

• Move up the funnel. Measure the main CTA click, the checkout start or a price view. The base is much larger (20%, 30%, 40%), so the required sample collapses. The risk is optimizing a metric that does not pull revenue, so pick a micro conversion you already know correlates with the purchase.
• Test big changes. Whole page against whole page, offer against offer. A large effect is the only kind that shows up with a small sample.
• Merge similar pages. If ten product pages share a template, test the change on the template and pool the traffic of all ten.
• Extend the window carefully. Twelve weeks is still defensible; six months is not, because your site, your pricing and your traffic mix change midway and contaminate the comparison.
• Use a Bayesian read. It does not create information that is not there, but it gives a more useful answer for deciding on thin data: the probability that the variation is better and the expected loss if you are wrong. See the Bayesian A/B test calculator.
• Pair it with qualitative research. Session recordings, heatmaps, five interviews and one usability test are cheap and surface obvious problems no A/B test needed to prove.

The full strategy lives in our guide on CRO for low traffic sites, and the underlying sample math is covered in what A/B testing is.

Limits of this calculator

The math assumes a binary metric (converted or not), stable traffic across the window, a balanced split between variations and visitors independent from one another. It does not model strong seasonality, campaigns that change your traffic mix mid test, novelty effect, or the time between a decision and building the next variation. The smallest detectable lift uses the normal approximation with variance evaluated at the base rate, the market convention for MDE: with a very low rate and a very small sample, treat the output as an order of magnitude rather than an exact figure.

FAQ

Can you run A/B tests with low traffic?

You can, but only for big effects. With low traffic your window holds a small sample, and a small sample can only tell large differences apart from noise. With 8,000 visitors per month, a 2% conversion rate and a 6 week window, the smallest detectable lift is around +37% relative: detail changes (button color, microcopy) will never show up, but a new offer or a new value proposition might. The calculator shows that ceiling for your own numbers.

What is the minimum traffic for an A/B test?

There is no universal number, because what matters is the combination of traffic, conversion rate and effect size. A practical reference: to detect a +10% relative lift on a 2% base within 6 weeks you would need roughly 115,000 visitors per month on the tested flow. If your base rate is higher (5% or 10%), the same relative lift costs far less sample. Run the calculator before assuming testing is off the table.

What should I do when there is not enough traffic to test?

Move up the funnel and change the size of the bet. Measure a more frequent micro conversion (main CTA click, checkout start, price view), which has a much larger base and fills the sample far sooner. Test radical changes instead of details. Merge similar pages into a single test. Use qualitative research, session recordings and heuristics for what you cannot prove quantitatively. And resist running an underpowered test: it almost always ends inconclusive.

Can I drop confidence from 95% to 90% just to be able to test?

You can, and the rigor selector lets you simulate it. Dropping to 90% confidence and 70% power cuts the required sample considerably, at the cost of more false positives and more false negatives. It is a defensible trade in exploratory work, when shipping the winning variation is cheap and you accept revisiting the decision later. It is not defensible when the decision is expensive or hard to reverse.

How is this different from the A/B test duration calculator?

The duration calculator starts from an effect you already chose and answers how many days it takes. This one starts from the window you are willing to wait and answers what fits inside it: the smallest visible lift, whether the lift you want makes it, and how much traffic is missing. It is the diagnostic tool for teams who suspect their site is too small to test, with the verdict and the plan B.

Keep going

If the verdict was positive, plan the test with the sample size calculator and call the result with the statistical significance calculator. If it was negative, start from what A/B testing is to understand what statistics can and cannot prove with the traffic you have today.