How to Actually Analyze an Incrementality Test Result
Incrementality analysis turns lift test output into budget decisions. Learn to read iROAS, relative lift, confidence intervals, and delayed conversions with a worked example.
On this page
- The four numbers that carry the analysis
- A worked example (hypothetical)
- The interval matters more than the point estimate
- Delayed conversions and where you cut the clock
- Reading a null result correctly
- Turning one result into a system
- When is the lift reading itself the mistake?
- Where the numbers in this analysis come from
- Sources
A finished incrementality test is not a decision. It is a table of numbers with uncertainty attached to every one of them. The analysis is where you decide whether the lift is real, whether it is precise enough to trust, and what it means for the next budget cycle. Plenty of teams run clean experiments and then misread the output, usually by fixating on the wrong metric or ignoring the confidence interval entirely.
This piece is about that reading step. If you need the concept first, start with what incrementality is and come back.
The four numbers that carry the analysis
Google's Conversion Lift reports several metrics, and each answers a different question. Per Google Ads, the core ones are:
- Incremental conversions (absolute lift): treatment conversions minus control conversions. The raw count of conversions that would not have happened without the ads.
- Relative lift: incremental conversions divided by control conversions, expressed as a percentage.
- Incremental cost per action (iCPA): total ad spend divided by incremental conversions.
- Incremental ROAS (iROAS): incremental conversion value divided by total ad spend.
The trap is treating relative lift as the headline. It scales between -1 and infinity, and Google itself warns that a brand with low control conversions can post a huge relative lift for a modest real effect. Two studies with 900% and 40% relative lift can describe the same underlying reality. Judge campaigns on absolute lift, iCPA, and iROAS, because those tie directly to money.
A worked example (hypothetical)
Assume a mid-size retailer runs a user-based Conversion Lift study on a Demand Gen campaign. Numbers are illustrative.
| Metric | Treatment | Control |
|---|---|---|
| Conversion value | $180,000 | $120,000 |
| Conversions | 3,000 | 2,400 |
| Ad spend | $30,000 | (held out) |
Work it through using the Google formulas:
- Incremental conversions = 3,000 - 2,400 = 600
- Incremental conversion value = $180,000 - $120,000 = $60,000
- iCPA = $30,000 / 600 = $50
- iROAS = $60,000 / $30,000 = 2.0
Now the standard Google Ads report for this campaign might show 2,200 attributed conversions and a reported ROAS near 6. The lift study says only 600 of those were caused by the ads. That gap is the whole point of running the analysis: attribution counts touchpoints, incrementality counts causes. If you want the mechanics of that difference, our attribution primer lays it out.
An iROAS of 2.0 clears break-even for most margins. So the campaign stays. If it had come back at 0.7, the honest read is that the platform was taking credit for demand that existed anyway.
The interval matters more than the point estimate
Every lift number is a midpoint inside a range. Google reports credible or confidence intervals precisely because the point estimate alone is misleading. In the 2025 measurement updates, Google noted advertisers can now choose a custom test size and view results at their preferred confidence level directly in the UI, per Google Ads.
Here is why it matters. Suppose that iROAS of 2.0 carries a 90% interval of 0.8 to 3.2. The midpoint says profitable. The lower bound says possibly underwater. Incremental conversion value tends to be more volatile than conversion counts, because a single bulk order can swing it, which widens intervals further. When the interval straddles your break-even line, you do not have an actionable result yet. You have a signal to extend the study, increase test size, or accept a lower confidence threshold with eyes open.
A disciplined analysis always reports the estimate and the interval together, then sizes the budget move to the pessimistic end of the range.
Delayed conversions and where you cut the clock
Some conversions land after the study window closes, even when the ad exposure happened inside it. Google's user-based Conversion Lift models these using observed conversion lag and, for Demand Gen studies, includes delayed incremental conversions by default, per Google Ads. For a product with a long consideration cycle, cutting analysis off at the study end date will understate true incremental value. Check whether your report includes projected conversions before you compare iROAS across channels with different sales cycles.
Reading a null result correctly
The hardest analysis is the one that comes back flat. A confidence interval that spans zero does not prove the channel does nothing. It means the test lacked the power to distinguish the effect from noise, or the effect is genuinely small. Amplitude points out that most user-level tests need adequate group sizes to detect a given lift, and undersized groups produce non-significant results even when the campaign works, per Amplitude.
So when you see a null, ask two questions before concluding anything. Was the minimum detectable effect smaller than the lift you could realistically expect? And was the control group large enough? If both answers are no, the correct read is inconclusive, and the fix is a bigger or longer test. Our guide to running a quarterly geo-holdout test covers the sizing math for the geo case.
Turning one result into a system
A single lift number is a snapshot at one point in time. The value compounds when you use it to correct your other models. Google frames incrementality, Marketing Mix Models, and attribution as a combined toolkit, where experimental results calibrate the models so their ongoing numbers track reality more closely, per Google Ads.
That calibration is the practical payoff. If your platform reports a ROAS of 6 and your lift study says the true incremental ROAS is 2, the ratio between them becomes a correction factor you apply to future platform numbers between tests. We walk through that arithmetic in the multiplier method, which keeps your dashboards honest without running a fresh experiment every week.
Google also lowered the entry point sharply. What once cost upward of $100,000 for a single experiment can now be run for around $5,000, and improved methodology returns conclusive results up to 50% more often, per Google Ads. That changes the cadence question. When a test is cheap and more likely to resolve, you can afford to re-measure your biggest channels quarterly and keep the correction factors current.
If your reported ROAS and your gut disagree, our Attribution Doctor can help pressure-test where the credit is leaking before you commit to a full study, and our data and analytics team runs the design and analysis end to end. The reading is where the money moves, so treat it with the same rigor you gave the test setup.
When is the lift reading itself the mistake?
Reading the output correctly is not the same as reading the right output. The situations where a clean study still produces the wrong decision:
-
Treating a platform lift study as a channel verdict. Conversion Lift compares users the platform held out against users it exposed, inside its own audience. It cannot see whether those incremental conversions came out of another channel's pocket. In practice you will see a healthy iROAS while blended revenue per marketing dollar does not move. Cross-check every platform study against a model-free ratio such as MER, and confirm with a geo test before the number moves a large budget.
-
Scaling on an average. An iROAS of 2.0 describes the spend level that was tested. The counterintuitive part is that the next dollar almost always returns less, so a "profitable" channel can be unprofitable at the margin. The trade-off: a scale test costs another study, but a budget increase decided on the average is a guess.
-
Value lift in a lumpy business. Where a few large orders dominate revenue, incremental conversion value swings on one deal that happened to land in treatment. The pitfall is a wide interval hiding behind a strong-looking point estimate. Read the conversion-count lift first and treat the value figure as supporting evidence.
-
Studies that straddle a promotion. A test window covering a sale measures the ads plus the offer, unless the holdout saw the same promotion through other channels. Either way the result does not transfer to a normal week.
That said, the mistake most teams make first is subtler: cutting a channel on a null result the study was never powered to detect. The downside is losing a working channel to noise.
Where the numbers in this analysis come from
Every figure here is drawn from the source cited beside it: the metric definitions and the relative-lift warning from Google Ads Help, the delayed-conversion handling and the cost and conclusiveness claims from Google's 2025 measurement update, and the power argument from Amplitude. The worked example is hypothetical; its treatment and control figures were chosen to make the arithmetic legible and are not a benchmark for any channel. Google's cost figures describe its own Conversion Lift product, so they say nothing about what a geo holdout you run yourself will cost. On engagements the platform definitions serve as shared vocabulary in the reconciliation meeting; the numbers that move budget come from the client's own test, reported with their interval. In our work with Cirque du Soleil on Dralion, closed-loop attribution across paid and organic let budget follow proven performance: the six-week run of 22 performances sold out at a 15x return on ad spend.