What's the Best Incrementality Testing Platform for You?

The best incrementality testing platform is the one that can run the test your decision needs, on your own sales data, and show you enough to judge whether the result holds up. There's no single incrementality number. "What is my current Meta spend doing?" and "What happens if I add 50%?" are different questions, and a test that answers one doesn't answer the other.

Below are the four questions ecommerce brands ask us most. Sometimes the right answer is "your data can't answer this yet," and we'll say when.

How do you choose the best incrementality testing platform?

Work through six questions in order. What decision are you making: cut, keep, launch, or scale? What effect do you need to know? What test can estimate that effect? Can your data support it? What could make the result wrong? Does the platform show you enough to check?

Check the minimum detectable effect before you launch. For a given design and test window, it tells you roughly how big an effect needs to be for the test to have a reasonable chance of seeing it. If the effect you care about is well below that, redesign the test before you spend on it. More in why a flat result isn't the same as zero.

Meta takes most of my budget. How much is incremental?

Run an inverse holdout. Turn Meta off or cut it in a set of regions, keep it running everywhere else, and compare actual sales in the holdout regions to an estimate of what they'd have done with no change. The difference estimates the incremental effect of the spend you removed during the test.

Platform reporting grades its own homework. Attribution answers a different question from incrementality, and research comparing observational estimates with randomized ad experiments has found the two can diverge substantially (Gordon et al., 2019). A geo test starts from your business outcome rather than attributed conversions, then estimates what that outcome would have been without the change.

Set it up. Choose a full pause or a partial cut. A pause estimates the effect of all current spend in those regions. A partial cut estimates the slice you removed, which sits closer to the margin. Pick regions that have moved together historically, and measure orders or revenue by ship-to region from your store. You need daily regional sales history long enough to build a reliable comparison.

What can throw it off. Excluding regions can push the same budget into the regions still running, which contaminates your comparison. Cut total budget roughly in proportion and check regional spend after launch to confirm it held.

Then there are confounders: a regional promo, a shipping change, weather, or a competitor's push that hits test and control regions unevenly. Matching regions on history won't catch a shock that starts after launch, and the estimate will credit or blame your spend for it. Log promos, stock-outs, and shipping changes by region, and watch the control regions for unexplained jumps. Then use a model that adjusts for known confounders and shows whether the result holds after adjustment. Adjust for things that would have happened anyway, not things the test itself could have caused.

Past exposure can keep working after spend changes, and conversions can lag, so set the test and post-test windows before launch.

What it tells you. Whether the removed spend earned back more than your breakeven. If you judge on same-period contribution profit, breakeven iROAS is roughly 1 divided by your contribution margin. At a 40% margin, that's 1 ÷ 0.40 = 2.5x. If you're optimizing to customer lifetime value, the threshold changes. Either way, it says nothing about the next dollar, or about half or double the budget.

At Stella, a feasibility check comes first: the minimum investment to hold out against current spend, correlation between candidate regions, and placebo tests that check whether the design finds spurious effects in untreated historical periods. Results are graded against a weighted synthetic control. Your team turns the regions off unless you ask us to.

I want to try linear TV, but people can't click it

Use a geo holdout. Air TV in test markets, keep comparable control markets dark, and compare sales by region against a synthetic control built from the pre-launch period. Last-click gives TV zero credit by design, and the TV vendor's own report is one more number you can't check.

Set it up. Buy TV you can target by market, like local spot by DMA. A national buy has nowhere to hold out. A weighted synthetic control blends several untreated markets to reproduce how the test markets moved before launch, instead of leaning on one hand-picked match. If that pre-launch fit is weak, fix market selection before you buy airtime. Good pre-launch fit still isn't enough on its own. The control markets have to stay a credible comparison during the flight.

Measure every material sales channel you can see by region, like your site, Amazon, and retail. Purchases often land in a different channel from the one that carried the ad. If a channel is missing, call the result lift in the channels you observed.

What can throw it off. Signals cross market borders, viewers travel, and national cable or streaming can reach your control markets. Keep buffers between test and control where you can, and log anything else that changes in one region during the flight. Those are confounders, and the model needs them as inputs to separate them from the effect of TV. TV builds and fades slowly, so set a post-flight window before launch. Fewer markets means a bigger minimum detectable effect.

What it tells you. Whether TV at this weight, creative, and market mix cleared breakeven, and whether a bigger second test is worth it. A three-market test doesn't predict a national rollout.

At Stella, we handle market selection and the synthetic control, then read lift on your observed regional sales.

I'm testing influencers or affiliates and can't create a geo holdout

Causal impact can work here, but only when the setup supports it. It forecasts what sales would have done without the campaign, using control series with a strong pre-launch relationship to your sales, then compares that forecast to what happened (Brodersen et al., 2015). Having no holdout doesn't make it valid by default.

Set it up. You need one clean launch date, like a big creator drop, and enough history before it to cover weekly swings and seasonality. You need control series that tracked your sales closely before launch, with a credible argument the campaign couldn't move them. An unfeatured product line isn't automatically safe, since a creator can lift your whole brand. Run placebo analyses over untreated periods. If the method keeps finding effects when nothing happened, the real estimate deserves much less trust. Longer walkthrough: causal impact analysis.

When to walk away. Always-on affiliate programs rarely have a clean start date. A creator post that lands the same week as a sitewide sale can't be separated from the sale. If your controls moved with the campaign, or stopped tracking your sales before launch, the estimate is biased. Then you can't tell yet, and that's the right answer. Stagger launches by time or region, or buy creators whose audience clusters in specific markets, and you've built a comparison for next time. In the meantime, promo codes and post-purchase surveys are useful directional evidence, but they aren't causal lift estimates.

What it tells you. Whether one campaign likely moved sales enough to repeat. Of the four designs, it leans on the strongest assumptions, because no comparison group is deliberately kept untreated while the campaign runs. Treat it as an estimate. At Stella, we run causal impact when the setup holds and tell you when it doesn't.

Is it safe to double my Meta budget?

A holdout on today's spend can't tell you whether doubling pays. To estimate the return on more spend, test more spend. Raise Meta in test regions while comparison regions remain at their planned baseline spend. Incremental revenue divided by incremental spend estimates the return on the increase you tested.

Your current iROAS is an average across all the spend you have now. It doesn't show the slope of the curve above today's budget, and extra spend often buys less as delivery reaches less responsive people. More in average vs. marginal iROAS.

Set it up. Make the increase big enough to read. A 25% bump usually isn't: the confidence interval on the extra revenue can come out so wide that you can't tell a 25% increase apart from no increase at all. A small spend change makes a small signal, and it gets lost in the normal day-to-day swings in regional sales. A proper scale test raises spend by 100% or 200% in test regions, so the change in revenue is large enough to estimate marginal ROAS with a useful range. Before launch, check that the test regions can absorb that much spend without frequency spiking or budget going unspent.

Better still, read three spend levels: today's baseline, +100%, and +200%. Three points let you sketch the marginal growth curve and see where returns start to flatten into saturation. Pinning down the exact saturation point from three noisy reads still takes an assumed curve shape, which is where an MMM calibrated with these results comes in. Run test regions in separate geo-targeted campaigns, and check that spend landed where you planned.

What can throw it off. A big jump changes more than spend. Learning can reset, and frequency and creative fatigue climb, so part of what you measure is the jump itself. Plan a ramp period instead of shrinking the test. Confounders matter here too: a promo or regional shock in only the test regions can look like a return on the extra spend. If you plan to treat relearning as a ramp, define that window before launch and report both the full period and the steady-state read. Do the same for any post-test window. Extending the window after you see the result is how a test gets talked into a number.

What it tells you. Whether each tested increase cleared breakeven, and roughly where on the curve returns start to flatten. Past the highest level you tested, you're extrapolating, so retest before you push further.

At Stella, we size the increases against the minimum detectable effect, run the test regions, adjust for confounders, and report marginal ROAS at each level against your breakeven.

What about Meta's or Google's own lift tests?

For some questions they're the cleaner option. Meta's Conversion Lift and Google's user-based Conversion Lift randomly withhold ads from a share of people and compare conversions between the groups. Randomizing people is stronger than matching regions, so if your question is "what is Meta doing for the conversions Meta can see," a platform lift study is a good place to start.

The limits are about scope. A platform lift study measures one platform, so it can't compare Meta against TV or tell you how to split budget. It counts conversions through the platform's own tracking, so Amazon, retail, and offline sales are usually missing unless you upload them. And the company selling the ads runs the test. Use one for a single-platform read, and use a geo test when the question is about your whole business outcome or more than one channel.

What changes for lead gen and B2B?

The designs carry over, but you measure a different outcome: qualified pipeline or closed revenue from your CRM, tied to the lead's region. Form fills are easy to inflate, so use a leading indicator like SQLs only after confirming it tracks closed revenue in your own history. Long sales cycles stretch the window, and lower volume makes a lift harder to detect. Most of Stella's work is DTC, but B2B brands with enough conversion volume can run these tests with us.

Does a higher price mean better measurement?

No. Price isn't evidence of rigor. Judge the design, the assumptions, the diagnostics, the uncertainty, and whether the result can support the decision you're making. An expensive engagement can still hand you in-sample fit and one iROAS with no range.

Across 225 Stella geo tests from a self-selected group of DTC brands, stronger pre-period fit was more closely associated with reaching significance than budget or duration were. That's descriptive, since brands, channels, and designs differ across the sample. Part of it is mechanical, too: closer fit means a narrower interval. That's the point. More budget or a longer test can't buy back a comparison that doesn't track.

Stella publishes its pricing: a free tier, self-serve plans from $750 to $6,000 a month, and managed engagements scoped per brand (as of October 2026). Holdout markets come from a location selection that includes placebo tests, results are graded against a weighted synthetic control, and the MMM is Bayesian and calibrated with holdout results. Every holdout result shows pre-period fit, error, and an uncertainty interval, so you can inspect the evidence behind the estimate.

What should the best incrementality platform show you?

Enough to judge how uncertain the estimate is (precision), whether the comparison plausibly estimates the causal effect (validity), and whether reasonable modeling choices would change the conclusion (robustness). A narrow interval doesn't rescue a bad comparison.

Two things to ask about every geo test. First, separation: changing spend in test markets shouldn't change exposure or sales in the controls, but travel, cross-border media, national promos, budget reallocation, and shared auctions can all leak. A good design reduces those risks and tells you which ones it couldn't remove. Second, assignment: most geo tests pick markets for historical fit rather than assigning them at random, which makes the quality of the comparison matter even more. And ask how the interval is built. For synthetic control, it usually comes from placebo runs on untreated regions or periods, not a textbook formula.

Frequently asked questions

What is the best incrementality testing platform?

The one that can run the test design your decision needs, on your own sales data, and show pre-period fit, minimum detectable effect, an uncertainty interval, and sensitivity checks on every result. There's no single best platform for every brand.

Can I run an incrementality test without a geo holdout?

Sometimes, with causal impact. It needs a clean launch date, enough pre-launch history, and control series the campaign couldn't plausibly affect. Without those, use a staggered or regional design instead.

How long should an incrementality test run?

Long enough to detect the effect you care about and catch delayed effects. There's no universal number. A pre-test power check sets it from your regional noise and the size of the change, and the window gets fixed before launch.

Should I use Meta's Conversion Lift instead of a geo test?

Use it when you only need Meta's effect on the conversions Meta can track. It randomizes people, which is a clean design. Use a geo test when you need your full business outcome, including Amazon or retail, or want to compare channels.

Is a cheaper incrementality tool less accurate?

Not necessarily. Accuracy comes from test design and diagnostics. Ask any vendor, at any price, for pre-period fit, the minimum detectable effect, an uncertainty interval, and sensitivity checks.

Not sure which test fits your question?

Bring your channel mix and the decision you need to make. We'll tell you which test fits, whether your data can support it, and whether Stella is the right partner. Book a demo.