A client showed me a Meta Conversion Lift study last month claiming their campaigns drove 312% incrementality. Three hundred and twelve percent. Every dollar spent created $3.12 of conversions that wouldn’t have happened otherwise.

I asked them what they thought that number meant. They said it proved their ads were working. I asked them what they’d do if the number had come back at 0%. They said they’d probably cut the budget.

That’s the trap. The Conversion Lift report is treated as ground truth, and ground truth doesn’t lie. Except platform-run lift tests aren’t ground truth. They’re a measurement product designed by the same company selling you the ads. There’s nothing inherently dishonest about that—it’s just worth being clear-eyed about what you’re actually getting.

What These Tests Actually Measure

Meta’s Conversion Lift, Google’s Brand Lift and Conversion Lift, TikTok’s Lift Studies—they all share a basic architecture. The platform divides eligible users into a test group (sees your ads) and a control group (doesn’t). At the end of the study period, the platform measures conversion rates in both groups, calls the difference incrementality, and reports it back to you.

In theory, this is the gold standard of measurement. Randomized controlled trial, exposed versus unexposed, clean comparison. In practice, there are at least four problems with this setup that don’t get discussed enough.

The first is that “eligible users” is the platform’s definition. The control group isn’t a random sample of the population—it’s a random sample of users the platform’s targeting algorithm decided were good prospects for your ads. These people are already disproportionately likely to convert. The lift you measure is incrementality relative to a high-propensity baseline, not absolute incrementality.

The second is that the conversion measurement on the control side is often platform-attributed. If a control-group user converts on your site, the platform counts it through the same pixel, conversion API, or modeling that counts test-group conversions. Any systematic measurement error applies to both groups, but the test group is more exposed to it because they actually saw the ads.

The third is that the platform decides when to call the test concluded. Tests that show negative lift have a habit of being extended. Tests showing strong lift get reported promptly. I can’t prove this is happening systematically, but I’ve observed enough cases to be suspicious.

The fourth is that the platforms charge you to run these tests. Meta requires a minimum spend commitment to access Conversion Lift. The product is, structurally, a value-added service for advertisers who already spend a lot. Whose incentive is served by reporting low lift to your largest customers?

What an Independent Test Looks Like

If you actually want to measure incrementality, the test needs to be designed by you and the holdout has to be properly random across your full target audience, not just the platform’s eligible pool.

The cleanest way I’ve seen this done is geo-based holdout testing. You pick matched pairs of geographic regions, turn off the campaign in one of each pair, leave it on in the other, and measure outcomes from your own data, not the platform’s. This requires enough volume that you can detect a meaningful effect, which usually means at least mid-six-figure spend and a few weeks of test duration.

For lower-volume advertisers, audience-based holdouts can work but require careful design. The key is that the holdout has to be defined and randomized outside the platform’s targeting system. Otherwise you’re back to measuring lift against high-propensity baselines.

Meta’s documentation is reasonably honest about the limitations of their lift product if you read it carefully. The marketing materials are less honest. Treat the documentation as the truth and the marketing as marketing.

What I Actually Believe About Platform Lift Numbers

The platform-reported numbers aren’t useless. They’re directional. If Meta says your campaign showed strong lift, it probably did at least some lift. If they say it showed no lift, that’s a serious red flag.

But the magnitudes are nearly always inflated relative to what an independent test would show. My rough heuristic, based on running both kinds of tests for the same campaigns: divide the platform-reported lift by somewhere between two and four to get closer to the truth. A 312% Meta-reported lift probably looks more like 70-150% under proper measurement, which is still a great result. Just not three-times-better-than-a-great-result.

For a category where I’ve seen this most starkly, ecommerce performance campaigns targeting existing customers, the gap is even wider. The platforms count repeat purchase activity that would have happened anyway as incremental. Independent holdout tests usually show that retargeting and lookalike-of-customer campaigns have far lower true incrementality than platform reports suggest.

How to Talk About This With Stakeholders

The hardest part of this isn’t the measurement work. It’s explaining to a CMO that the report they’ve been showing the board for the last three quarters has been overstating reality.

What I’ve found works: don’t attack the platform numbers directly. Run an independent test in parallel and present both results without commentary. Let the gap speak for itself. The conversation will go where it needs to go.

If you preempt it by framing platform measurement as flawed before you have data, you’ll lose the political fight. If you let the data do the framing, you usually win. Marketers respect data. They don’t respect being told their data is wrong without proof.

I’ve also found it helps to focus the conversation on decisions rather than on which number is “right.” The question isn’t whether Meta’s lift number is accurate. The question is whether the marketing budget allocation would change if we believed the independent number. Usually it would, sometimes substantially. That’s the framing that gets attention.

Where This Is Going

The platforms know this is a problem. They’re investing in better measurement products—incrementality APIs, more transparent holdout designs, third-party verification programs. Some of these are genuine improvements. Some are window-dressing.

The honest path forward for sophisticated advertisers is to maintain measurement independence. Run your own incrementality tests on a recurring basis. Use platform-reported numbers as one signal among several, not as ground truth. And be skeptical of any measurement product where the seller is also the buyer’s largest reporting source.

If you want help designing an incrementality testing program that holds up to scrutiny, the kind of custom AI development work that combines marketing data engineering with proper statistical design is increasingly important—and increasingly rare to do well in-house. Either way, don’t accept a 312% lift number at face value.