Calculate iROAS by dividing estimated incremental revenue by the ad spend used to generate it. In a randomized holdout test with unequal group sizes, first scale the control group's outcome to the size of the test group. Subtracting raw conversion totals would give the wrong answer.
For the definition and comparison with ordinary ROAS, see what is iROAS. This guide focuses on the calculation. The free iROAS calculator now accepts randomized test and control revenue as well as an existing incremental estimate. Download the editable Excel example and test plan to follow each calculation.
Use group revenue when order values differ
For randomized individuals with comparable follow-up, use revenue per assigned person directly:
Control revenue per person = control revenue / assigned control population
Expected test revenue without treatment = control revenue per person × assigned test population
Estimated incremental revenue = observed test revenue − expected test revenue without treatment
The free workbook uses 10,000 test participants with $33,600 revenue and 2,000 control participants with $5,600 revenue. Expected baseline revenue is $28,000, leaving an estimated $5,600 increment. These are fictional numbers. They avoid the extra assumption that each group has the same order value.
Retain zero-revenue participants in the denominator. Do not select only people who saw an ad, clicked or purchased. Group totals do not reveal outcome variance, so they cannot by themselves produce a confidence interval. A regional experiment needs a method that accounts for its assignment unit and market differences.
The holdout-test formula
Test conversion rate = test conversions / test audience
Control conversion rate = control conversions / control audience
Incremental conversions = (test rate − control rate) × test audience
Incremental revenue ≈ incremental conversions × average order value
iROAS = incremental revenue / test ad spend
Use assigned test and control groups, measured over the same period, rather than selecting only people who actually saw or clicked an ad. Post-assignment selection can bias the estimate. For a user-level test, define a conversion as a converted user when calculating a user conversion rate.
The average-order-value shortcut assumes a suitable common order value. If groups differ in order values or repeat purchases, compare revenue per assigned user instead:
Incremental revenue = test revenue − control revenue × (test audience / control audience)
A geo experiment needs an appropriate regional counterfactual model. Do not apply this simple randomized-user formula to unmatched regions with different baseline demand. See how to run a geo experiment.
Worked example: unequal group sizes
The following numbers are illustrative, not a client result. A randomized test assigns 10,000 people to advertising and 2,000 to a holdout. Each converted person makes one $80 order.
| Input | Test | Control |
|---|---|---|
| Assigned people | 10,000 | 2,000 |
| Converted people | 420 | 70 |
| Conversion rate | 4.2% | 3.5% |
| Revenue | $33,600 | $5,600 |
Test ad spend is $8,000. The expected baseline for 10,000 people is 70 × 5 = 350 conversions. Estimated incremental conversions are 420 − 350 = 70, not 420 − 70 = 350.
Incremental revenue = 70 × $80 = $5,600
iROAS = $5,600 / $8,000 = 0.70x
The absolute conversion-rate lift is 0.7 percentage points. The relative lift is 4.2% / 3.5% − 1 = 20%. The estimated incremental share of test conversions is 70 / 420 = 16.7%. Those are three different quantities.
If a platform separately attributes $33,600 to the same spend and scope, reported ROAS is 4.20x. The resulting calibration factor is $5,600 / $33,600 = 16.7%. Do not assume a platform automatically attributes every test-group order.
Copy the calculation into Excel or Google Sheets
Enter these illustrative inputs in column B and keep rates as decimals. The formulas work in Excel and Google Sheets without locale-dependent function separators.
| Cell | Label | Input or formula |
|---|---|---|
| B2 | Test audience | 10000 |
| B3 | Test conversions | 420 |
| B4 | Control audience | 2000 |
| B5 | Control conversions | 70 |
| B6 | Average order value | 80 |
| B7 | Test ad spend | 8000 |
| B8 | Contribution margin before ads | 60% |
| B10 | Test conversion rate | =B3/B2 |
| B11 | Control conversion rate | =B5/B4 |
| B12 | Incremental conversions | =(B10-B11)*B2 |
| B13 | Incremental revenue | =B12*B6 |
| B14 | iROAS | =B13/B7 |
| B15 | Break-even revenue iROAS | =1/B8 |
| B16 | Incremental contribution after ads | =B13*B8-B7 |
Audience sizes and ad spend must be positive; margin must be greater than zero for a finite break-even ratio. Negative incremental revenue is possible and should not be silently replaced with zero.
What is a good iROAS for your margin?
A revenue-based iROAS of 1x only recovers ad spend before product and other variable costs. If your contribution margin before advertising is 60%, break-even is 1 / 0.60 = 1.67x, assuming no additional marketing costs.
| Contribution margin before ads | Break-even revenue iROAS |
|---|---|
| 20% | 5.00x |
| 40% | 2.50x |
| 60% | 1.67x |
| 80% | 1.25x |
In the example, incremental contribution after ads is $5,600 × 60% − $8,000 = −$4,640. Include agency, creative or other relevant costs separately when assessing marketing ROI. A point estimate alone is not enough to decide whether to scale or stop a campaign.
Check uncertainty before acting
A test estimates a counterfactual; it does not identify exactly which customers were caused to buy. Review confidence or credible intervals, randomization quality, contamination, the planned analysis and whether the observation window captures the outcome you care about.
If the incremental revenue interval crosses zero, do not present the point estimate as a proven positive effect. If it crosses your profitability threshold, the profit decision remains uncertain even when lift is positive.
A calibration factor can help interpret subsequent attribution reports, but only while audience, spend, attribution window and conditions remain comparable. It is not a permanent correction, and a 20% relative lift is not automatically a 20% calibration factor.