How Meta Incremental Attribution Actually Works

Meta incremental attribution changes the question the system asks. Standard attribution asks whether a conversion followed an eligible click or view. Incremental attribution asks whether the ad made that conversion more likely. That is the better question. It is not the same as having a verified answer, because Meta has not published how the model is trained, what it predicts at, or how accurately it recovers real lift for any given advertiser.

Table of contents

This piece separates three things and keeps them separate: what Meta documents, what follows from what Meta documents, and what is one plausible explanation among several. Where the evidence stops, we say so rather than filling the gap.

Meta documents this	Meta has not disclosed this
Machine learning models predict whether a conversion was caused by an ad¹	What the models are trained on
The output can drive both reporting and delivery optimization²	Whether training data is pooled across advertisers
Reported results are "modeled"³	Whether account-specific features are used
Results are scoped to relative comparison between Meta campaigns³	Whether prediction is event, person, or campaign level
Reporting data does not exist before April 1, 2025³	How predictions are calibrated, or how far off they typically are
Conversion Lift is a separate product that uses randomized holdouts⁴	Any uncertainty range around the output

Meta Business Help Center: 1. Incremental attribution, 2. About attribution models, 3. Incremental attribution reporting, 4. About Conversion Lift. The right-hand column has no equivalent citation, which is the point of the table.

What is Meta incremental attribution?

It is an attribution model in Ads Manager, available for reporting and for delivery optimization. Standard attribution credits conversions that fall inside a click or view window. Incremental attribution tries to credit only conversions the model predicts the ad caused, which means fewer conversions get credited and the account gets optimized toward a different group of people.

The reason to want that is growth. Someone already about to buy is likely to produce a standard attributed conversion and be worth little incrementally, because the purchase was coming either way. That is revenue you already had, and paying to re-count it adds nothing. Incremental optimization is trying to tell the two cases apart, so the same budget chases revenue advertising creates rather than revenue it takes credit for. How well it manages that is the open question.

The reporting mode lives under Columns, Compare Attribution Models. It adds an incremental conversions column beside your standard numbers, changes nothing about delivery, and is not backfilled before April 1, 2025. The delivery mode is the real switch: select it under the performance goal at ad set setup and Meta optimizes toward conversions its models predict advertising caused, rather than toward people likely to convert after seeing an ad. Over the following weeks that changes who sees your ads and which creative wins.

One detail most coverage skips: attribution settings, the controls for click-through, view-through, engage-through and window length, exist for standard attribution only. Choose incremental and those controls disappear, so the definition of what counts moves entirely inside Meta's model.

Why are three different numbers all called conversions?

Most of the confusion about this feature comes from treating three different quantities as versions of the same number. They are not. They answer different questions and they are produced in different ways.

Attributed conversions. Conversions assigned to Meta under a set of rules. A qualifying ad interaction happened, a conversion followed, the model gave Meta credit. That establishes a sequence. It does not establish that the ad changed anything.

Predicted incremental conversions. Conversions, or portions of conversion credit, that a model estimates would not have happened without the advertising. The word predicted is doing a lot of work there.

Experimentally estimated incremental conversions. The average difference in outcomes between comparable treated and untreated groups. An experiment produces this. It tells you how many additional outcomes the advertising produced across a population.

No method observes whether one particular order would still have happened without the ad. You see one world, the one where the ad was available, never the same customer in the version of the week where it was not. That unseen version is the counterfactual. An experiment gets around it by comparing groups rather than individuals, which yields an average rather than a verdict on any specific purchase, and a model can spread that average across individual conversions, but those event-level labels stay predictions.

Conversion Lift is Meta's product in the third category, documented separately, with randomized test and holdout groups. Incremental attribution sits in the second category unless Meta discloses account-specific experimental machinery it has not described.

Does Meta run a holdout on your campaign?

Meta's public documentation does not say that enabling incremental attribution launches a randomized holdout on your account. That is the defensible version of the claim, and anything more certain outruns the evidence.

The setting's help page covers it in about sixty words: an attribution model that optimizes delivery for incremental conversions, using machine learning models that predict whether a conversion is caused by an ad. The attribution models page repeats it. The reporting page calls the output "modeled results" and scopes it to "the relative incrementality between Meta campaigns." None of the three says holdout, control group, experiment, or lift study.

So the reasonable reading is that these are model outputs rather than account-specific experimental estimates. Plenty of coverage describes Meta withholding ads from a slice of your audience when you switch the setting on. No Meta documentation we have found supports that, and most of it appears to be describing Conversion Lift.

What runs underneath remains an open question. It could be a treatment-effect model trained on historical lift studies, a calibration factor applied on top of standard attribution, or different methods for optimization and for reporting. The documentation does not distinguish between them.

Ghost ads are one way a platform could do this. The auction runs for both groups, and when the advertiser would have won an impression for a control user, the platform logs the opportunity and withholds the ad. That produces a control group closely resembling the people who would have been exposed. Johnson, Lewis and Nubbemeyer documented the method in the Journal of Marketing Research in 2017, and the paper is free on SSRN. It shows a platform can combine experimentation, auction data and prediction. It does not show that Meta incremental attribution works this way.

Why did your ROAS drop when you switched?

Incremental reporting may assign less conversion credit than standard reporting. When it does, credited volume falls and reported ROAS falls with it. That drop tells you the two models disagree about credit. It does not tell you the business got worse, and it does not tell you which number is closer to the truth.

Most tests end here. Someone flips it on, watches ROAS fall by a third, and flips it back.

Say standard reports 4.0x and incremental reports 2.6x. The 1.4x gap is model disagreement. It might reflect demand your ads were capturing rather than creating. It might reflect real differences in incrementality. It might also reflect calibration choices or plain model error. Ads Manager cannot separate those, and neither can we.

Seer Interactive ran a comparison in July 2025 across six accounts and about $1.05M of April spend. Meta reported 87% of conversions as incremental. Checked against GA4, the same accounts came out at 67%. That is two systems disagreeing, and it establishes nothing more. GA4 is not an experimental benchmark, so it cannot establish the correct incremental percentage either, and comparing across attribution models is the comparison Meta's own documentation warns against.

You will also see performance claims quoted around this feature: 20%, sometimes 46%, sometimes 24%. None appear in Meta's documentation. They circulate through agency posts citing each other, with no sample size or confidence interval attached.

Is Meta incremental attribution a measurement tool?

It is a modeled measurement and optimization system. Based on what Meta has published, its method is not disclosed and its accuracy for your account has not been independently validated. That is a narrower claim than saying it is not measurement.

The model may still be useful. A predicted-incrementality objective is better aligned with the question an advertiser cares about than a rule crediting every eligible post-ad conversion. Better alignment can improve delivery even when the reported score is imperfect. Alignment and accuracy are separate properties, and this feature has a strong claim on the first and an undisclosed one on the second.

There is also a validation problem worth naming carefully. Meta controls delivery, generates the predicted-incrementality signal, and reports the performance produced under it. When a platform-reported metric improves, two explanations fit: delivery produced more incremental outcomes, or the model assigned more incremental credit. From inside the platform those look the same. None of this shows Meta inflates anything, but it is a reason to check consequential budget decisions against something that does not depend on Meta's own modeled reporting.

Gordon et al., in Marketing Science in 2019, found that commonly used observational methods often failed to reproduce experimental effects across 15 large Facebook studies. That research does not test Meta's incremental attribution model. What it shows is that model sophistication does not substitute for validation against an experiment, which is exactly what the help pages reviewed here leave unaddressed.

How should you test it?

It is worth testing as an optimization objective, though on its own it is not evidence the measurement problem is solved. A defensible evaluation looks like this.

Define the decision before the test. Pick the business outcome that matters: revenue, orders, new customers, contribution margin. Then decide the smallest improvement that would change what you do. A detectable difference too small to act on is not a useful result.

Make the conditions comparable. Randomize campaigns or markets between standard and incremental optimization where you can. Hold the conversion event, eligibility rules, budget treatment, creative supply and dates as steady as the design allows. Comparing this month on one setting against last month on the other confounds the setting with everything else that moved.

Measure outside the columns being compared. The primary outcome should come from order, CRM or finance data, not from whichever attribution model is under test. A Conversion Lift study gives randomized people-level evidence where it is available. A geo experiment works when people-level randomization is impractical, though geo designs carry their own requirements: enough power, stable control markets, and care around spillover and structural change. We have written about when a geo test is the wrong tool.

Report an interval, not a winner. Give the estimated difference, the uncertainty around it, and whether the test could have detected the improvement your decision required. "The incremental campaign won" leaves out everything you need to act on.

Our incrementality testing guide goes further into designing the test itself.

How do you make platforms bid on incremental sales?

You change the number you send them. Ad platforms optimize toward the conversion value you report, which makes that value the objective function. Most brands report gross revenue, so the system is being asked to buy revenue, including the revenue that was arriving anyway. Sending an incrementality-adjusted value changes what it goes looking for.

Start with the mechanism. Google's value-based bidding strategies, Maximize conversion value and Target ROAS, bid toward the conversion values you report. Google's own documentation says those can be real economic values like revenue, or proxy values like a lead score. Meta and TikTok behave the same way for value optimization. The platform does not know what your numbers mean. It maximizes whatever you send it.

Send gross revenue and you have told the auction to find purchases. Send an incrementality-adjusted value and you have told it to find purchases that would not have happened otherwise. The machinery does not change, only the objective does.

The write paths are documented and unglamorous. Meta has the Conversions API. Google has offline conversion imports, plus conversion adjustments that restate the value of a conversion already on record, keyed on transaction ID or GCLID. TikTok has the Events API. Moving values from a warehouse into those endpoints on a schedule is reverse ETL.

Where Stella fits, stated precisely

A holdout does not reveal which individual orders were caused by advertising. Nothing does. What it produces is an estimate of aggregate lift for a tested population over a tested period, with an interval around it.

That estimate calibrates the value. The experiment anchors the expected incremental contribution, a model spreads it across conversions the experiment did not directly cover, and the next experiment checks whether the extrapolation still holds. Stella runs that loop and pushes the calibrated values back into Meta, Google and TikTok. The feature is called Push to Platform, and underneath it is reverse ETL.

Two limits are worth stating plainly. Feeding an experiment-calibrated value to a platform is not the same as handing it a list of conversions known individually to be incremental. No such list exists, ours included. Every conversion-level number in this design is still a prediction.

Ours does not escape modeling either. What differs is the anchor and where the values live. The anchor is an experiment you can inspect and run again. The values sit in your warehouse, so a conversion can be re-priced after the fact instead of carrying a score you are not allowed to see.

The idea is not new. Lewis and Wong made the case for unifying bidding, attribution and experimentation in Incrementality Bidding and Attribution in 2022, and Randall Lewis co-authored the ghost ads research above. What changed is that the write paths now exist on every major platform.

Book a scoping call to see how much of your Meta revenue is incremental, and what it would take to prove it.

Where does incremental attribution go next?

Every major platform is moving the same way. The objective is shifting from conversions that followed an ad to conversions the platform predicts the ad caused, and each platform will grade its own prediction with a model it does not publish. That turns the decision into a question about whose calibration you trust.

That changes what a measurement vendor is for. Reporting a second opinion next to the platform number matters less once the platform is already optimizing toward its own estimate of the same quantity. The number steering delivery is the one that counts.

The advantage goes to advertisers who own the value they send. A platform's internal estimate cannot be inspected, re-run, or corrected from the outside. A value you compute yourself can be, and it stays yours when the platform changes its model.

Expect the disclosure gap to stay open. Nothing obliges a platform to publish how it grades its own delivery.

Frequently asked questions

Does Meta run a holdout test on my campaign when I use incremental attribution?

Meta's documentation does not say that it does. It describes machine learning models that predict whether a conversion was caused by an ad, calls the reported output modeled, and documents Conversion Lift separately as the product that uses randomized holdouts. The reasonable reading is that incremental attribution produces model estimates rather than account-specific experimental ones.

Why are my incremental conversions lower than my standard conversions?

Because the incremental model assigns less credit than the standard model. Standard attribution counts everything inside the window, incremental counts a subset. The size of the gap tells you how much the two models disagree. It does not by itself establish how much demand your ads were capturing rather than creating, which takes an experiment to answer.

Is incremental attribution the same as a Meta Conversion Lift study?

No. A Conversion Lift study is a randomized experiment on your account over a defined window, with a control group withheld from your ads, and it produces an estimate with uncertainty attached. Incremental attribution is an always-on model whose training, calibration and error rates are not disclosed, and whose output carries no published interval.

Can I use Meta incremental attribution as my reporting KPI?

It targets a more decision-relevant quantity than standard attribution, which is a point in its favor. Its accuracy for your account is not publicly established, which is the problem with making it the number you report to finance. Meta produces that score and Meta is the party being scored.

How do I check whether incremental attribution actually improved my results?

Compare the two settings on an outcome measured outside both of them. Randomize between them where you can, hold everything else comparable, and read the result off order or finance data rather than off either attribution column. Comparing Meta's two settings against each other in Ads Manager will not answer it, and Meta's own documentation advises against that comparison.