Incrementality testing: how to know if your ads actually work

Every ad platform will tell you your campaign drove a pile of conversions. What none of them will tell you, on their own, is how many of those conversions would have happened anyway. Incrementality testing answers that question, and until recently it was priced out of reach for most small and mid-size advertisers. That changed when Google rebuilt its incrementality tools around Bayesian statistics and dropped the practical minimum spend for a test from around $100,000 to about $5,000. If you have never run one, this is the year it becomes realistic.

Attribution tells a story. Incrementality tests it.

Attribution, whether it is Google's data-driven model or Meta's own reporting, assigns credit to touchpoints along a path a customer took before converting. It is useful and it updates in real time, but it cannot answer a harder question: would that person have bought anyway, through organic search, a direct visit, or a competitor's ad they never clicked. A retargeting campaign can show a glowing ROAS in-platform while adding almost nothing, because it mostly reaches people who were already going to convert.

Incrementality testing removes that ambiguity by design. You hold back ads from a matched group, whether by geography, by user, or by a public service announcement served instead of your creative, and compare outcomes. The difference between the exposed group and the control group is your actual lift, closer to a clinical trial than a dashboard report, and that is exactly why it has been expensive and slow to run at scale.

What changed: Google's threshold dropped from six figures to five

Google's incrementality testing used to require enough spend and conversion volume that a frequentist statistical test could reach significance, which in practice meant guidance closer to $100,000 for a reliable read. In 2025 Google rebuilt the methodology around Bayesian modeling, which incorporates prior data and hierarchical structure rather than requiring a large fresh sample to hit a p-value threshold. According to coverage of the change, that shift lets Google produce a usable incrementality read from around $5,000 in test spend, with tests running anywhere from about 7 to 56 days and 28 days treated as the practical default.

That is a twenty-fold drop in the entry point. It matters because incrementality testing was previously something only advertisers spending six or seven figures a month could justify running per channel, per quarter. A business spending $8,000 a month on Search or Performance Max can now realistically run a test on a slice of that budget without derailing the rest of the account.

How the major platforms let you test lift

Google is not the only platform with a native tool. Meta's Conversion Lift and TikTok's Conversion Lift Study both use a similar structure: a portion of your matched audience is randomly withheld from seeing your ads, then post-test conversion rates are compared between the exposed and control groups. Both are free to run inside the ad account, though they need enough conversion volume during the test window, generally in the low hundreds of conversions, to produce a statistically meaningful gap rather than noise.

Geo holdout tests remain the most flexible option because they work on any channel, including ones without a native lift tool, and do not depend on user-level tracking that is increasingly unreliable anyway. You split your active markets into two matched groups, typically by state or metro cluster, pause spend in one group for the test window, then compare conversion rates. The tradeoff is that smaller advertisers need to run these longer, often four to six weeks instead of the two weeks a national brand might use, because lower weekly volume takes longer to produce a clean signal.

Why this is happening now, not two years ago

Two forces pushed platforms to build cheaper incrementality tools. First, signal loss. Third-party cookie restrictions, iOS tracking prompts, and browser-level privacy changes have degraded user-level attribution across the board, a shift we covered in our piece on signal loss and first-party data. When individual-level tracking gets noisier, aggregate causal methods like incrementality testing and marketing mix modeling become more trustworthy by comparison, not less.

Second, AI-driven bidding. Google Ads, Meta Advantage+ and similar automated systems now treat conversion data as an input that steers bidding, not just a report you read afterward. If that data is inflated because attribution over-credits ads for conversions that would have happened anyway, the algorithm optimizes toward a fiction and spends more to chase it. EMARKETER's coverage of measurement trends going into 2026 pointed to the same pattern: aggregate, privacy-resilient measurement is becoming the foundation layer, with incrementality testing and marketing mix modeling triangulated against platform attribution rather than any one method standing alone.

Running a first test without overcomplicating it

1. Pick one campaign or channel, ideally the one where you most suspect platform ROAS is inflated. Branded search and retargeting are common suspects.
2. Choose the matching method that fits your setup. If you sell in multiple states, a geo holdout is usually simplest. If you are single-market, use the platform's native lift tool.
3. Set the test window before you start and do not touch the campaign's budget, targeting or creative mid-test. Changing variables mid-flight invalidates the read.
4. Run a pre-test period of similar length to confirm your test and control groups actually track each other closely before you introduce the difference. If they diverge by more than a few percentage points before the test even starts, the groups are not properly matched.
5. Compare the lift number against what the platform's own attribution reported for that same spend and period. The gap between the two is the real headline.
6. Repeat per channel a few times a year rather than treating one test as permanent. Auction dynamics, audiences and creative all shift, and last year's lift number goes stale.

Where incrementality fits with everything else you already track

Incrementality testing is not a replacement for daily attribution reporting, and it is not a replacement for marketing mix modeling either. It sits between them. Attribution is fast and granular but biased toward over-crediting ads. MMM is slower and gives a portfolio-level view across online and offline channels but cannot isolate a single campaign the way a controlled test can. Incrementality testing is the periodic, causal check that recalibrates the other two. Feed the results into how you read platform dashboards day to day, and you get a measurement stack that catches itself when a channel's reported performance drifts from reality, which matters most when you are trying to scale ad spend without quietly inflating your real customer acquisition cost.

One data point worth keeping in view: the ANA's Q1 2026 Programmatic Transparency Benchmark found that top-performing advertisers converted 54.0 percent of their programmatic spend into fraud-free, viewable, measurable impressions, against 32.1 percent for the lowest-performing cohort, a 21.9 point gap the benchmark called its widest on record. Cleaner delivery and honest measurement tend to travel together, so this is not a project to run on top of an unmonitored account.

Frequently asked questions

What is incrementality testing in advertising?

Incrementality testing measures the sales, leads or conversions an ad campaign actually caused, by comparing a group exposed to the ads against a similar group that was not. It answers a different question than attribution: not who gets credit for a conversion, but whether the ad made that conversion happen at all.

How much does incrementality testing cost?

Google's Bayesian incrementality tests now run from roughly $5,000 in ad spend, down from a prior guideline of around $100,000, using probabilistic modeling that needs less data to reach a confident read. Meta and TikTok's built-in lift tests are free to run inside the ad account but need enough conversion volume, generally a few hundred conversions during the test, to produce a reliable result.

Is incrementality testing better than attribution?

They answer different questions and both have value. Attribution tells you which touchpoints a converting customer passed through, which is useful for daily optimization. Incrementality tells you whether the spend caused the outcome, which is what matters when deciding whether to keep, cut or scale a channel. Most measurement teams now use both alongside marketing mix modeling rather than picking one.

Can a small business run an incrementality test?

Yes, with realistic expectations. A business spending $2,000 to $10,000 a month on one channel can run a geo holdout or a platform-native lift test, but should expect to run it for four to six weeks rather than the two weeks a large advertiser might use, since lower weekly volume needs more time to produce a statistically reliable read.

What is a geo holdout test?

A geo holdout splits your markets into two matched groups, typically by state, metro area or zip code cluster. One group keeps seeing ads, the other has spend paused entirely for the test window. Comparing conversion rates between the two groups after the test shows how much of your result depended on the ads running there.

Does incrementality testing replace attribution reporting in ad platforms?

No. Platform attribution reports still drive day-to-day bid and budget decisions because they update constantly and cover every campaign. Incrementality tests are periodic checks, run a few times a year per channel, that recalibrate whether the platform's reported numbers are close to reality or badly overstated.

The takeaway

For years, incrementality testing was a large-advertiser privilege, because doing it properly needed a budget most small and mid-size businesses could not spare for a single experiment. The Bayesian shift at Google, plus free native tools at Meta and TikTok, has quietly moved that threshold within reach of a business spending a few thousand dollars a month per channel. You do not need to overhaul your whole measurement setup this quarter. Pick the one campaign you trust the least, run one test, and let the gap between the platform's number and the real number tell you where your next budget conversation should actually start.