An RPV calculator divides total revenue by total visitors to give revenue per visitor (RPV), and its significance test checks whether the gap in RPV between your control and variant is bigger than random noise.
Because revenue per visitor is an average of a skewed number (most visitors spend $0), you need the standard deviation of revenue per visitor, not just conversion counts, and you compare the two means with a t-test.
Revenue per visitor calculator: what it calculates
The calculator above gives you three numbers for each variant: revenue per visitor, conversion rate and average order value (AOV).
It also gives the relative lift in RPV between variant and control.
The formulas are simple:
- RPV = total revenue / visitors
- Conversion rate = orders / visitors
- AOV = total revenue / orders
- RPV = conversion rate × AOV
- RPV lift = (RPV variant - RPV control) / RPV control
That last identity is the reason RPV is the main metric on most ecommerce tests.
A change can win on conversion and lose on order value, or the reverse, and RPV nets the two.
Our guide to revenue per visitor as a test metric covers when to pick it over conversion rate.
For the parts on their own, use the conversion rate calculator or the AOV calculator.
Two tests from our database show why the split matters.
When Elevate A/B Testing moved the second product photo into the main slot for The Cornish Company, RPV rose from £3.32 to £3.60 and conversion rose from 11.66% to 12.61%.
By our math, AOV stayed flat at about £28.50, so the gain came from more buyers.
When Elevate A/B Testing listed the biggest bundle first for Repeat Undies, RPV went from €3.54 to €3.84 (+8.47%) while conversion only moved from 5.6% to 5.74%.
Our math puts AOV at about €63.21 versus €66.90, so most of the lift came from bigger orders.
Neither source published sample sizes or significance, so treat both as directional.
Revenue significance calculator: inputs, formula and worked example
A conversion-rate significance test only needs visitors and orders.
A revenue test needs more, because two variants with the same RPV can have very different spreads. Enter for each variant:
- Visitors (everyone assigned, including non-buyers)
- Total revenue
- Standard deviation of revenue per visitor, with non-buyers counted as $0. Most testing tools and a per-visitor export from your analytics give you this.
The calculator runs Welch's t-test (Welch, Biometrika, 1947), which does not assume both groups have the same variance:
- Difference: d = RPV B - RPV A
- Standard error: SE = sqrt(SD_A² / n_A + SD_B² / n_B)
- Test statistic: t = d / SE
- 95% confidence interval for the difference: d ± 1.96 × SE (with thousands of visitors per arm, t is close to normal)
Here is a worked example with made-up but realistic numbers. Each arm gets 20,000 visitors.
| Control | Variant | |
|---|---|---|
| Visitors | 20,000 | 20,000 |
| Orders | 600 | 640 |
| Revenue | $36,000 | $40,320 |
| Conversion rate | 3.0% | 3.2% |
| AOV | $60.00 | $63.00 |
| RPV | $1.80 | $2.016 |
| SD of revenue per visitor | $12.87 | $13.70 |
The variant shows a 12.0% RPV lift. The difference is $0.216 and SE = sqrt(12.87² / 20,000 + 13.70² / 20,000) = $0.133. That gives t = 1.63 and a two-sided p-value of about 0.10.
The 95% interval runs from -$0.04 to +$0.48 per visitor.
That range includes zero, so this test has not shown a real difference yet, even though the lift looks big.
Notice that each SD is about seven times the RPV.
That is normal for revenue, and it is why revenue tests need more traffic.
To detect a $0.216 difference with 80% power at 95% confidence, you would need about 59,000 visitors per arm: n = (1.96 + 0.84)² × (SD_A² + SD_B²) / d².
The sample size calculator does the same for conversion rates.
Ronny Kohavi and co-authors give a similar example in their 2007 practical guide to controlled experiments: a site where 5% of users buy, average spend is $3.75 per user and the SD is $30 needs over 1.6 million users to detect a 5% revenue change at 90% power.
How to read the result
- p-value below your threshold (usually 0.05): the RPV gap is unlikely to be noise alone. It does not tell you the size of the effect.
- Look at the interval, not just the verdict. An interval of +$0.01 to +$0.40 is a win, but the true effect could be tiny. Read our guide to confidence intervals.
- Check which part moved. If RPV won but conversion fell, the result depends on a few larger orders and may not repeat.
Large losses happen too. Milos's test that changed a 2-pack offer to "Buy 2 Get 1 Free" at the same price cut RPV by 52.38%.
Dave Diederen's test that added a countdown timer for a European market cut it by 35.61%.
Dave Diederen also removed the free-shipping threshold and saw RPV rise 9.17%, a result you could easily miss if you only tracked conversion.
@DaveDiederen
@elevate_ab@elevate_ab@DaveDiederen
@CROguy_Common mistakes
- Running a conversion-rate test on revenue. A two-proportion z-test ignores order value. Use the A/B test significance calculator for conversions and this one for RPV. Not sure which applies? See which statistical test to use.
- Leaving out non-buyers. If you only count purchasers, you are measuring AOV, and the SD you enter will be wrong.
- Letting one huge order decide the test. A single wholesale order can flip the result. Optimizely's outlier smoothing flags revenue values above the mean plus three standard deviations and replaces them with the harmonic mean. You have to switch it on per experiment, and only accounts with access to the feature see that option. Whatever rule you use, set it before the test starts.
- Mixing sessions and users. Assignment is usually per user, so divide by users. Dividing by sessions makes the SD formula assume independence that isn't there.
- Stopping early on a big lift. Revenue swings a lot from day to day. Decide your sample size first, or use CUPED to cut variance.
- Ignoring refunds and currencies. Use net revenue if returns differ by variant, and convert to one currency before you calculate.
For more revenue-based results, browse the full test database or start with the A/B testing guide.