A geo test readout typically leads with a headline in the shape of "+5% sales lift, +$350k incremental revenue", and sometimes a single incremental ROAS figure like "$1.30 per $1". The lift is almost always going to be there, because we spent money so obviously sales increased. The bigger problem is the uncertainty around these numbers, which rarely gets talked about. In our experience, for most brands under ~$300M a year, the range on incremental ROAS is too wide to make a budget call on, so there are two questions worth asking every time, one before the test and one after.
Here is what that looked like on a test we scoped for a DTC brand.
The headline vs the range
In this case the iROAS came out at ~$1.30, which sounds like a profitable channel, but it is only the middle of a range that runs from roughly $0 to $3. That range takes in "this channel loses money" and "this channel should get double the budget", with ~1 in 3 odds the spend actually lost money. So even the iROAS number is misleading when it's reported on its own, because the decision depends on where the whole range sits, not where its middle is.
The test designer's own planning tool told the same story before anything launched. Power, the chance the test catches an effect that's really there, was ~54% with a 10% holdout, ~57% at 30% and ~59% at 50%, and it barely moved between 2, 3 and 4 weeks. Getting to ~68% would have meant roughly doubling the channel's spend for a month. The brand would have given up holdout revenue for a result that couldn't change the decision, so we didn't run it.
A geo test can prove a channel works and still not tell you whether to spend more on it.
Two questions to ask before you pay for a geo test
Before the test: what's the power? Under ~65%, a real effect gets missed 1 time in 3 or more, and I wouldn't bother (80% is the textbook minimum). If adding weeks doesn't move it, which it usually won't, the only real fixes are a bigger spend change or testing a bigger channel.
After the test: what's the 90% range on iROAS? Not the lift, and not a single iROAS number. If the range reaches below breakeven, or across whatever threshold you'd need to scale, the test hasn't answered the question. A good test designer can also tell you the range a design is expected to produce before anything goes dark, which is the cheapest time to find out it'll be too wide.
A breakeven channel and a profitable one can read the same
A single geo test is one draw from a wide spread of possible results, and at most DTC sizes the spread for a channel that breaks even overlaps heavily with the spread for one that's worth scaling. We simulated a common design (a $500k-a-week brand, a channel at 10% of revenue, a 4-week holdout) on two channels, one that truly returns $1.00 per dollar and one that truly returns $1.50, and looked at the spread of results a single test could report for each.
So a readout of $1.30 doesn't tell you which channel you have. Going on the single number, you'd scale the breakeven channel ~2 times in 10 and pass on the profitable one ~3 times in 10, which is why the range matters more than the headline. Running 8 weeks instead of 4 only narrows the spread by ~25%, because regional demand shocks (weather, local press, a competitor's promo) carry over week to week, and a bigger brand helps by about the same before it flattens out. Doubling the spend change halves it, but for a channel at 10% of revenue going dark is already the biggest cut you can make, so the only way further is doubling its budget in the test regions, which is usually a cost nobody wants to pay for a test.
What we did instead
We put the budget into reading several imperfect measures side by side, every 2 weeks, and treated disagreements between them as findings. None of them is precise on its own, but they fail in different ways, so when three of four point the same direction you can act with more confidence than one wide test range gives you.
A first-party quality yardstick. Cost per first-time quality visitor, meaning a new visitor who views 2+ pages including a product page, from a first-party pixel that saw ~20% more sessions than GA4. It moves within days of a spend change, long before revenue settles.
Marketing mix modeling. Reconciled against actuals every couple of months, with weekly error held to ~3%, so it covers the cross-channel and carryover effects a short test misses.
First-party multi-touch attribution. Directional at best for credit, but useful for seeing which journeys a channel actually shows up in.
A post-purchase survey. Biased toward memorable channels, which is useful in itself when it disagrees with the other three.
Reading these together is what let the team cut four long-tail channels that made up ~17% of spend, with each cut backed by more than one measure. The full story is in our measurement reset case study.
A geo testing vendor's own study lands here too
Recast, which sells geo lift testing as well as marketing mix modeling, published a simulation that shows the same tradeoff. Across 32,000 simulated geo tests with a known 7.5% lift, one treated market against 20 controls and a 15-day read, the most cautious method missed the real lift ~9 times in 10, and the most sensitive declared a winner ~30% of the time when there was no lift at all. Their design is harsher than a typical DTC test, so 4 to 8 week multi-region tests do better, but the direction matches ours, and Lewis and Rao found the median ROI interval across the advertising experiments they studied was over 100 percentage points wide.
When a geo test is worth running
I would run one when the channel is roughly 15-20% of revenue or more, its spend can move a lot in the test regions, the power is at least ~65% and the expected range clears your threshold. A lot means going dark or doubling, not a 10% nudge. If those can't all be true, the test costs you the holdout revenue and returns a range you could have predicted in planning.
If the plan can't be filled in honestly, that's usually the answer, and reading several measures side by side is the better use of the budget. Happy to walk through how we'd approach it for your channel mix.
Common questions
Are geo tests worth it for DTC brands?
A geo test is typically worth running when the channel is roughly 15-20% of revenue or more, you can change its spend a lot in the test regions, the test has at least ~65% power and the expected range on incremental ROAS clears your decision threshold. For most brands under ~$300M a year testing a mid-size channel, the range is too wide to separate scaling from cutting, even when the test reports a significant lift.
What questions should I ask a geo testing vendor?
Before the test, ask for the power and the 90% range on incremental ROAS the design is expected to produce. After the test, ask for the 90% range on incremental ROAS rather than the lift or a single iROAS number. If the power is under ~65%, or the range reaches below breakeven, the test is unlikely to change a budget decision.
What power should a geo test have?
80% is the textbook minimum. Below ~65%, a real effect gets missed 1 time in 3 or more, which we think makes the test not worth the holdout revenue. If power doesn't improve with more weeks, a bigger spend change or a bigger channel is usually the only real fix.
What is the difference between lift and incremental ROAS in a geo test?
Lift tells you whether sales rose in the test regions, and with real spend behind a channel they almost always will. Incremental ROAS tells you how much revenue each extra dollar of spend returned, which is what a budget decision needs. In a test we scoped, a ~5% lift and ~$350k of incremental revenue translated to ~$1.30 per $1 with a range from roughly $0 to $3.
How long should a geo test run?
In our simulation, doubling a test from 4 to 8 weeks narrowed the incremental ROAS range by only ~25%, while doubling the size of the spend change halved it. If the range crosses your decision threshold, change the spend change or the channel you test rather than adding weeks.
What should a DTC brand use instead of a geo test?
For channels too small to test well, read several imperfect measures side by side on a fixed cadence: a first-party quality metric such as cost per first-time quality visitor, a marketing mix model reconciled against actuals, first-party multi-touch attribution and a post-purchase survey. Each is biased in a different way, so agreement across three or four is stronger evidence than one wide test range, and disagreement tells you where to look.
We size and read geo tests as part of a measurement rebuild, including the power and expected range before anything goes dark, and the measures to read alongside when a test can't answer the question.
See how it works →The two-channel chart is simulated with known true incremental ROAS of 1.0 and 1.5 (no client data), assuming perfect seasonality control, no carryover and no leakage into holdout regions, so real tests typically read wider. Figures for the scoped test are rounded and partly illustrative, and its power estimates come from the test designer's own planning tool.