A/B Testing With Low Traffic: How Much Traffic You Actually Need
How much traffic you need for an A/B test, why conversions matter more than visitors, the minimum sample size by lift, how long a test takes at your traffic level, and what to do when you do not have the volume.
The short answer: As of October 8, 2026, the honest answer to "how much traffic do you need for A/B testing" is that traffic is the wrong question. The gate is conversions. To detect a 20% relative lift at 95% significance and 80% power, you need about 400 conversions per version, whether your page converts at 2% or 10%. A 10% lift needs about 1,600 per version. If your page produces 100 conversions a month, test big changes, pool pages, and use heatmaps and research for everything else.
This guide is for founders and small teams who want to A/B test a site that gets thousands of visitors a month, not millions. It covers the math, how long a test will take at your volume, and what to do when the answer is "too long."
TL;DR
- Conversions per version needed ≈ 15.68 ÷ lift². About 392 for a 20% lift, 1,568 for 10%, 2,788 for 7.5%.
- Your conversion rate barely changes that count. It changes how many visitors it takes to collect it.
- At 3,000 visitors a month and a 5% conversion rate, detecting a 20% lift takes about 24 weeks. Detecting a 10% lift takes over a year.
- Low traffic means testing bigger changes, pooling similar pages, running for a fixed planned duration and not peeking.
- Check feasibility first with the sample size calculator and the test duration calculator.
Why conversions, not visitors, decide whether you can test
Most advice on A/B testing low traffic sets a visitor threshold, such as 10,000 visitors a month. That is the wrong variable.
A test is trying to tell a real difference from random noise. The noise comes from how many conversions you observe. A page with 10,000 visitors at 2% produces 200 conversions a month. A page with 10,000 visitors at 5% produces 500. Same traffic, very different testing power.
So the number to plan around is conversions per month on the page you want to test, and the smallest lift you care about.
"Lift" here is always relative. A 20% lift on a 5% conversion rate means 5% becomes 6%.
The formula: minimum sample size for an A/B test
For a two-version test with a two-sided, two-proportion z-test at 95% significance (z = 1.96) and 80% power (z = 0.84), the visitors you need per version are:
Visitors per version = 2 × p̄ × (1 - p̄) × (1.96 + 0.84)² ÷ (p₂ - p₁)²
where p₁ is your current conversion rate, p₂ is p₁ × (1 + lift), and p̄ is their average. Multiply by p₁ to get conversions per version.
When the conversion rate is small, almost everything in that formula cancels except the lift. What is left is a rule of thumb you can do in your head:
Conversions needed per version ≈ 15.68 ÷ lift² (lift as a decimal, so 20% is 0.2)
The 15.68 is 2 × (1.96 + 0.84)². The table below shows the shortcut next to the exact formula at five baseline rates.
| Lift to detect | Shortcut (15.68 ÷ lift²) | At 1% baseline | At 2% | At 3% | At 5% | At 10% | Visitors per version at 3% |
|---|---|---|---|---|---|---|---|
| 5% | 6,272 | 6,363 | 6,297 | 6,231 | 6,099 | 5,770 | 207,704 |
| 7.5% | 2,788 | 2,862 | 2,832 | 2,802 | 2,742 | 2,592 | 93,403 |
| 10% | 1,568 | 1,629 | 1,612 | 1,595 | 1,560 | 1,474 | 53,152 |
| 15% | 697 | 741 | 733 | 725 | 709 | 669 | 24,167 |
| 20% | 392 | 426 | 422 | 417 | 408 | 384 | 13,900 |
| 30% | 174 | 198 | 196 | 193 | 189 | 177 | 6,449 |
| 50% | 63 | 77 | 76 | 75 | 74 | 69 | 2,516 |
All figures are conversions per version unless labeled otherwise. A two-version test needs twice this in total.
What the table shows:
- The conversion count is nearly flat across baseline rates. A 20% lift needs 384 to 426 conversions per version whether the page converts at 1% or 10%.
- Halving the lift roughly quadruples the requirement. Going from a 20% lift to a 10% lift takes you from about 400 to about 1,600 conversions per version.
- Small lifts are where most real wins live. Optimizely reports that about 12% of experiments produce a significant win on the primary metric (Optimizely, across 127,000 experiments). Among tests that did win, the median lift was 7.5% (Analytics Toolkit, 1,001 tests, 2022). Detecting a 7.5% lift takes about 2,800 conversions per version.
That is the core tension of A/B testing with low traffic. The lifts that are common are the ones that need the most data. For an established program, a realistic lift to plan for is 5% to 10% (Convert; Analytics Toolkit median planned MDE 6%). A 20% or larger lift is a setting for redesigns and bold changes, not button tweaks. For more on picking the number, see the minimum detectable effect guide, and for what the significance and power settings mean, see A/B testing statistical significance explained.
How long an A/B test takes at your traffic level
Duration is the question that matters on a low-traffic site. The table shows weeks to finish a two-version test with traffic split evenly, at five monthly traffic levels and two conversion rates.
| Monthly visitors | Conversion rate | Conversions per month | 10% lift | 20% lift | 30% lift |
|---|---|---|---|---|---|
| 1,000 | 2% | 20 | Over 2 years | Over 2 years | 86 weeks |
| 1,000 | 5% | 50 | Over 2 years | 71 weeks | 33 weeks |
| 3,000 | 2% | 60 | Over 2 years | 62 weeks | 29 weeks |
| 3,000 | 5% | 150 | 91 weeks | 24 weeks | 11 weeks |
| 10,000 | 2% | 200 | 71 weeks | 19 weeks | 9 weeks |
| 10,000 | 5% | 500 | 28 weeks | 8 weeks | 4 weeks |
| 30,000 | 2% | 600 | 24 weeks | 7 weeks | 3 weeks |
| 30,000 | 5% | 1,500 | 10 weeks | 3 weeks | 2 weeks |
| 100,000 | 2% | 2,000 | 8 weeks | 2 weeks | 2 weeks (minimum) |
| 100,000 | 5% | 5,000 | 3 weeks | 2 weeks (minimum) | 2 weeks (minimum) |
"2 weeks (minimum)" means the math finishes sooner, but you should still run for two full weeks so every day of the week is counted twice (CXL; Optimizely and VWO set a one-week floor).
If you know your monthly conversions but not your traffic, this shorter table is enough. It is computed at a 3% baseline; other rates land within a few percent.
| Conversions per month on the tested page | 10% lift | 20% lift | 30% lift |
|---|---|---|---|
| 50 | Over 2 years | 73 weeks | 34 weeks |
| 100 | Over 2 years | 37 weeks | 17 weeks |
| 250 | 56 weeks | 15 weeks | 7 weeks |
| 500 | 28 weeks | 8 weeks | 4 weeks |
| 1,000 | 14 weeks | 4 weeks | 2 weeks |
| 2,500 | 6 weeks | 2 weeks | 2 weeks (minimum) |
CXL and Ton Wesseling put the point where a recurring testing program becomes practical at 1,000 conversions a month (CXL / Ton Wesseling). Below that, you can still test, but each test has to earn its slot.
Methodology. Every number in this post comes from src/lib/ab-test-math.ts, the same code behind our sample size calculator and test duration calculator. Visitors per version use the formula above, rounded up. Conversions per version = visitors per version × baseline rate, rounded. Duration assumes two versions with an even split: daily visitors = monthly visitors × 12 ÷ 365, days = visitors per version ÷ (daily visitors ÷ 2), rounded up, then converted to whole weeks, rounded up. "Over 2 years" means more than 104 weeks. The conversion rates are planning inputs, not benchmarks. Put your own rate into the calculator.
What to do when you don't have the volume
A long number in the duration table does not mean you can't improve the page. It means you pick tests your traffic can answer, and use other methods for the rest.
1. Test bigger changes
The fastest way to shorten a test is to raise the lift you are trying to detect. That means testing changes that could plausibly move conversions by 20% or more: a new headline that states a different promise, a different offer or price presentation, a much shorter signup form, or a rebuilt page tested as a split URL test.
Button colors and small copy edits rarely move results that much. On a page with 150 conversions a month, a 30% lift takes about 11 weeks to detect and a 10% lift takes about 91 weeks.
The tradeoff: a bundled change tells you the new version is better, not which part made it better. On a low-traffic site, that is usually the right trade.
2. Pool similar pages into one test
If several pages share a template and an audience, apply the same change to all of them and count conversions across the group.
At 3,000 visitors a month and a 5% conversion rate, one page needs about 24 weeks to detect a 20% lift. Three similar pages pooled to 9,000 visitors a month at the same rate need about 8 weeks.
Only pool pages where the visitor, the offer and the change are alike. Pooling a pricing page with a blog post gives you an average that describes neither.
3. Use a higher-funnel metric, with care
Clicks on the main call to action happen far more often than purchases or signups. At a 20% click rate, detecting a 20% lift takes about 336 clicks per version, which at 3,000 visitors a month is about 5 weeks. The same lift measured on a 5% signup rate takes about 24 weeks.
The catch is that a click is not the outcome you care about. A version can win on clicks and lose on signups, for example by making a vague promise that people click on and then abandon. Use the higher-funnel metric to decide faster, then check that the final conversion did not fall for the winning version before you keep it.
4. Run tests one after another, not at the same time
Running two tests on the same page splits the same small pool of visitors. Run one test at a time on each page, to its planned sample size, then start the next. A test queue ranked by expected impact beats several half-powered tests running in parallel.
5. Use qualitative research and heatmaps for the small stuff
A test is one way to learn. It is not the only one. For a low-traffic site, these are often faster:
- Heatmaps and scroll maps show where visitors stop reading, what they click that is not a link, and whether they reach the call to action at all. See how to use heatmaps to improve website conversions.
- Five to ten customer interviews surface the objection your page never answers.
- User tests where someone tries to complete the signup while thinking aloud expose broken steps in minutes.
If research finds something clearly broken, such as a form error or a missing price, fix it. You don't need a test to prove that a broken step is bad. Compare the before and after period, and note what else changed.
6. Decide the sample size up front and don't peek
Set the planned sample size and the end date before launch. Then leave the test alone until it gets there.
Checking results daily and stopping the first time the result looks significant inflates false positives well beyond the 5% you planned for. Low-traffic tests are the most exposed to this, because their early results swing the most. If the planned duration is longer than you can wait, the fix is a bigger change, not an earlier stop.
Common mistakes when A/B testing with low traffic
| Mistake | Why it hurts | What to do instead |
|---|---|---|
| Stopping when the result first looks significant | Early results on small samples swing widely, so many early "winners" are noise | Stop at the planned sample size, after at least two full weeks |
| Testing three or four versions at once | Every version needs its own full sample, so each one you add stretches the test | Test one challenger against the control |
| Testing tiny tweaks | Small lifts need thousands of conversions per version | Test changes that could plausibly move results by 20% or more |
| Planning around visitors | Two pages with the same traffic can have very different conversion counts | Plan around conversions per month on the tested page |
| Calling an inconclusive test a tie | No significant difference is not proof of no difference; the test may simply lack power | Record it as inconclusive and note the lift it could have detected |
| Changing the page or traffic mix mid-test | A new ad campaign or a page edit changes who is in the test | Freeze the page and note any campaign changes |
On extra versions, the cost is direct. At 10,000 visitors a month and a 5% conversion rate, a test to detect a 20% lift takes about 8 weeks with two versions, about 11 weeks with three and about 15 weeks with four. That is before any correction for comparing several versions against the control, which pushes it further.
How to check whether your test is feasible before you launch
Spend five minutes on this before you build anything.
- Find the page's conversions per month. Use the page you will test and the goal you will count, not sitewide totals.
- Pick the smallest lift worth acting on. Be honest. If only a 25% lift would change your decision, plan for 25%.
- Enter both into the sample size calculator. It returns the visitors and conversions each version needs.
- Enter your traffic into the test duration calculator. It returns how many days the test will run.
- Decide. If the duration is acceptable, launch and set the end date. If it is not, go back to the previous section: make the change bigger, pool pages, or switch to research.
If you are a solo founder choosing a stack for this, the best CRO tools for solo founders covers the options by budget and traffic.
How Humblytics fits
Humblytics runs A/B tests on your live site with a no-code visual editor, alongside analytics, heatmaps and funnels in one script. Here is what matters for a low-traffic site, and what does not change.
- Tests can stop automatically after a set duration (30 days by default). Significance is shown in real time against a 95% confidence threshold, but reaching it does not end the test: you decide when to call it. Significance is calculated in real time with a two-proportion test, the same family of test used for the numbers above. Promoting the winner stays a human decision.
- Set the duration from the calculator on a low-volume page. If your page needs 24 weeks, a 30-day limit will end the test before it has the sample. Use the planned duration as the stop rule.
- Heatmaps sit next to the test, so the research step above uses the same tool.
- Tests can be scored on a conversion or on revenue from a connected payment processor.
- Pricing is public. Business is $79 a month and allows 5 active A/B tests. Scale is $279 a month with unlimited tests. There is a 14-day free trial and no free tier. See pricing.
Humblytics does not make a small page converge faster. No tool can; the math above applies to every testing product. If your page cannot resolve a test in a reasonable time, the analytics, funnels and heatmaps are still useful from day one, and testing can wait until volume or the size of the change makes it worthwhile. See A/B testing in Humblytics.
Sources and freshness
- Humblytics,
src/lib/ab-test-math.ts, the formula behind the sample size calculator and test duration calculator. Supports every sample size, conversion count and duration in this post. Benchmarks quoted from the same file with their stated sources: Optimizely, across 127,000 experiments (12% win rate); Analytics Toolkit, 1,001 tests, 2022 (7.5% median winning lift); CXL / Ton Wesseling (1,000 conversions a month for a practical program); CXL; Optimizely and VWO set a one-week floor (two-week minimum duration); Convert; Analytics Toolkit median planned MDE 6% (realistic lift range of 5% to 10%). - Humblytics product facts (duration-based auto-stop, 30-day default duration, two-proportion significance, revenue scoring, human promotion of winners) from Humblytics' published product information, current as of October 8, 2026.
- Humblytics plan prices and limits from the Humblytics pricing page, current as of October 8, 2026.
The conversion rates in the tables are planning inputs, not industry benchmarks. Plan prices and limits can change; check the pricing page before you buy.