# How much traffic do you need for A/B testing?

> The honest answer has three inputs and no marketing spin. Here's how to calculate it for your own site, and what to do when the number comes back too big.

- Source: https://bixlow.com/en/blog/how-much-traffic-do-you-need-ab-testing
- Language: en
- Published: 2026-09-30T09:00:00.000Z
- Updated: 2026-10-06T09:28:18.400Z
- Author: fernando
- Topics: sample size, A/B testing, statistical significance, low traffic, minimum detectable effect, new brands, CRO
- Publisher: Bixlow

---
"How much traffic do I need?" is the first question every brand asks before starting with experimentation, and it's usually answered badly — either with a scary round number like "100,000 visits a month" or with a reassuring "any traffic works!". Neither is true, because the answer isn't a number. It's a function of three things you control.

## The three inputs

**Your baseline conversion rate.** The lower it is, the more sessions you need. Converting at 0.8% is dramatically more expensive to test than converting at 6%, because rare events need bigger samples to distinguish from chance.

**The minimum detectable effect.** How big an improvement you need to be able to see. This is the input people forget, and it's the one with the most leverage: halving the effect you want to detect roughly quadruples the sample you need.

**Your confidence and power.** Conventionally 95% confidence and 80% power. Confidence is your tolerance for calling a winner that isn't real; power is your tolerance for missing a real winner.

Notice what isn't on that list: the number of variants only matters because it divides your traffic, and the tool you use doesn't matter at all. Anyone selling you a platform that "needs less traffic" is selling you a lower confidence threshold with a nicer name.

## Some real numbers

At 95% confidence and 80% power, sessions needed **per variant**:

| Baseline | Detect +10% | Detect +20% | Detect +50% |
|---|---|---|---|
| 1% conversion | ~62,000 | ~16,000 | ~3,000 |
| 2% conversion | ~31,000 | ~8,000 | ~1,400 |
| 5% conversion | ~12,000 | ~3,100 | ~550 |
| 10% conversion | ~5,600 | ~1,450 | ~260 |

Read it as a menu, not as a verdict. A brand with 2% conversion and 500 sessions a day on its product pages has 250 per variant per day. Detecting +20% takes about a month. Detecting +10% takes four months. Detecting +50% takes under a week.

Bixlow runs this calculation from your actual data when you configure an experiment — it reads your baseline from your own traffic rather than asking you to guess it, and shows the expected duration before you launch. If the number is absurd you find out in the configuration screen, not eight weeks later.

## The lever nobody uses: pick a better metric

The table above is about purchase conversion. But you don't always have to measure purchase.

If your test changes something on the product page, add-to-cart happens maybe eight times more often than purchase. Measuring add-to-cart as your primary metric cuts your required sample by roughly the same factor. The trade-off is real — add-to-cart lifts don't always survive to revenue — which is why you pair it with a guardrail on revenue per session.

This is a legitimate technique when used honestly: measure the closest event to the change you made, and guard the event you actually care about. It becomes dishonest when the guardrail is missing and the add-to-cart number goes in the report as if it were revenue.

## When the number really is too big

Sometimes the honest answer is that you cannot test this hypothesis. That is information, not failure. Three things to do with it:

**Test it upstream.** You can't detect a 5% checkout improvement, but you can probably detect whether a completely different landing page for your Meta traffic works, because that effect is large.

**Test on a bigger population.** If the hypothesis only applies to one segment, the segment may be too small. Sometimes the same idea applies site-wide and becomes measurable.

**Decide without testing.** Not every decision needs an experiment. If the change is cheap, reversible and directionally obvious, ship it. Reserve your sample for the decisions where you genuinely don't know the answer and the cost of being wrong is high.

## What about sequential testing?

You'll see tools advertising "peek any time" sequential methods. They're real statistics, not a trick, and they do let you stop early when an effect is very large. What they don't do is make small effects detectable on small samples. They change *when* you can stop, not *how much* signal exists in your data.

Bixlow uses fixed-horizon tests with a declared sample size, and simply does not show a winner until the evidence is there. It's the more conservative choice, and for teams without a statistician on staff it's the one that produces fewer wrong decisions.

## Key takeaways
- The answer depends on your baseline conversion, the effect you want to detect and your confidence level — not on your tool.
- Halving the effect you want to detect roughly quadruples the sample required.
- Measuring a more frequent event (add-to-cart instead of purchase) cuts the sample you need, but only if you keep a revenue guardrail.
- "You can't test this" is a valid, useful answer. Spend your sample where you genuinely don't know.

---

## Meet our AI CRO agent

What you just read, put to work by an agent: it understands your visitors, decides what to test or personalize, builds it on your site without a developer and tells you when a result is real.

https://bixlow.com/en/ai-cro-agent