Incrementality testing: how to find out which sales your ads really caused

Incrementality is the share of your results that would not have happened without the ads. If 1,000 people bought and 850 of them would have bought anyway, your ads were responsible for 150 purchases. Attribution answers a different question: which touchpoint gets credit for each of the 1,000 purchases. Every platform attributes generously to itself, and none of them knows what would have happened without its ads.

That gap is not academic. A connected-car app that steers its ads with our agent system compared platform numbers with its own revenue. In Germany in September 2026, Meta, Google and TikTok together claimed 2.2 times the revenue the company actually made; Google alone reported €31.3k while the company made €23.4k. Attribution overlap explains part of that. Incrementality testing answers the next question: how much of the real €23.4k did the ads cause?

After twelve years of running ad accounts, agency side and client side, I trust one kind of evidence above dashboards: what happens to sales when the ads are off somewhere. If you want the basics of credit assignment first, read our guide to marketing attribution models.

Lift % = (conversions with ads − expected conversions without ads) ÷ expected conversions without ads

Incremental ROAS (iROAS) = incremental revenue ÷ ad spend

Incrementality factor = incremental conversions ÷ conversions the platform attributed

There are two families. Experiments split people or regions into a group with ads and a group without. Models estimate the effect from historical spend and sales. Experiments give cleaner answers for one channel at one time; models cover all channels but need calibration.

Enough that the effect you expect is clearly larger than the normal ups and downs of your sales. The published minimums give a floor.

• Google Conversion Lift based on users requires $5,000 and at least 1,000 observed conversions. Tests can be as short as 7 days, Google recommends more than 14, especially with long conversion lag (Google Ads Help).
• Google's geo-based Conversion Lift is available through an account representative, the campaigns must target a single country, and Google recommends the budget after a feasibility check (Google Ads Help).
• Meta runs Conversion Lift self-serve for advertisers who meet its criteria. As a guide, the ad account needs a campaign that started in the past year with at least $5,000 spend and 500 conversions; longer campaigns are prorated, so 180 days need $10,000 and 1,000 conversions (Meta Business Help Center). For lift tests Meta treats a confidence of 90% or higher as statistically reliable (Meta Business Help Center).
• GeoLift recommends daily data, at least 25 pre-treatment periods across 20 or more geo units, 4 to 5 times the test length in stable history (ideally 52 weeks), and a test of at least 15 days with daily data or 4 to 6 weeks with weekly data. The test should cover at least one full purchase cycle (GeoLift Best Practices).

Meta itself recommends people-based tests such as Conversion Lift where possible, because they have higher statistical power than geo tests (GeoLift). Geo tests win when a platform offers no lift study, when you want to test several channels at once, or when your real outcome, such as store visits or CRM deals, never reaches the platform.

1. Write down the question and the number you judge by. For example: "How many orders does Meta add in Germany?" measured in shop orders by delivery region, not in Meta's reported purchases.
2. Choose the geo unit. Use the most detailed level your platforms can target, such as cities or postal code areas. GeoLift works best with 20 or more units.
3. Pull daily orders per region for as far back as you can, ideally 52 weeks, and at least four to five times the planned test length.
4. Pick test regions whose sales moved in step with the rest in the past. GeoLift's power analysis does this for you and tells you the smallest effect you could detect.
5. Decide the treatment. A holdout switches ads off in the test regions. A lift test adds budget there instead. Holdouts are easier to read; lift tests cost less in lost sales.
6. Set the duration: at least one full purchase cycle and at least 15 days with daily data. Add a few days after the end for delayed conversions.
7. Freeze everything else: no national promotions, price changes or new TV during the test, and no campaign edits in the control regions.
8. Exclude the test regions in every campaign of the channel you test, then check delivery by region after the first day.
9. Do not stop early because the first week looks good or bad.
10. Compare actual orders in the test regions with what the control regions predicted, and calculate lift, iROAS and the incrementality factor.

Start with the confidence or interval, not the point estimate. A lift of 12% with an interval from 2% to 22% is a real effect with an uncertain size. A lift of 12% with an interval from minus 8% to 32% is inconclusive.

• Compare iROAS with your break-even ROAS. If ads cost €50 per incremental order and an order earns €40 gross margin, the channel loses money even if its attributed ROAS looks fine.
• Use the incrementality factor to discount platform numbers until the next test. It is a correction, not a law, and it drifts as campaigns, seasons and competition change.
• Treat an inconclusive result as "we could not measure it", not as "the ads do nothing". Rerun bigger or longer.
• Check other channels. If branded search rises in regions where Meta is off, some of Meta's effect moved there. Your total sales per region capture that; platform reports do not.

Test again after large changes in budget, offer or creative, and at least a few times a year for your largest channel. The numbers you steer on between tests should come from your own revenue, which we explain in our guide to true ROAS.

• Too few regions or too short a test, so normal noise hides the effect.
• Running the test during Black Friday, a price change or a national campaign that hits test and control regions differently.
• Spillover: people who live in a control region see ads at work in a test region, or ship to a different address than where they saw the ad.
• Editing campaigns in the middle of the test, which changes the treatment halfway.
• Measuring the result in platform conversions instead of your own orders or leads.
• Ignoring conversion lag, so late purchases fall outside the test window.
• Reading one test as permanent truth instead of a snapshot of one channel at one budget level.

Google's Meridian is an open-source marketing mix model that Google opened to all marketers in January 2025 (MarTech). Meta's Robyn is an open-source alternative in R and Python (GitHub). A model covers every channel at once, but its estimates can drift. Meridian therefore offers a way to turn results from incrementality experiments into priors for the model, with adjustments for spend, recency and test duration (Google for Developers). In practice: run a geo test on your largest channel, feed the result into the model, and let the model handle the budget split between tests.

Marketing AIgency is hosted software from CBX Media LLC. AI agents check ad accounts and steer on your company number, meaning real revenue or real leads from Shopify, your CRM, a spreadsheet or our first-party pixel, instead of platform ROAS. Every proposal goes into an approval queue with its reasoning. Meta is fully live; Google Ads and TikTok connect the same way with one-click login and run like Meta.

The Attribution add-on includes geo-lift support, next to a first-party pixel, a post-purchase survey and sending conversions back to Meta through the Conversions API. Geo tests run on Meta only for now. If you want to add a customer's own answer to the picture, our guide to the post-purchase survey shows how. Plans have fixed monthly prices with no share of ad spend (see pricing).