Ads Manager's standard attribution credits conversions that happened within a set window after someone saw, clicked or engaged with an ad. That is correlation, not cause. Meta's incremental attribution model goes further by predicting which conversions an ad caused, but it is still Meta's model, not a test of your account. To prove what Meta adds, compare people or regions that could see your ads with ones that could not. Conversion Lift does that with a randomized holdout, self-serve in Experiments for eligible accounts. Otherwise, a geo holdout read in your own sales data is the fallback, using GeoLift, an open-source, research-only tool from Meta Open Source. Robyn, Meta Marketing Science's experimental open-source MMM, then guides cross-channel budgets, calibrated by those tests. A/B tests pick the better strategy, but without a holdout they can't prove incrementality.
Attributed is not incremental
An attributed conversion is a bookkeeping rule: under standard attribution, Meta credits an off-ad action, such as a website purchase, to your ad if it happened within a set number of days after someone viewed or clicked it. Some of those buyers would have bought anyway, from repeat customers to people already searching for your brand. An incremental conversion is one the ads caused. Meta defines lift as the difference in conversions caused by the presence of your ads, and only a comparison with a group that couldn't see them measures it directly.
Meta now also offers incremental attribution as an attribution model at the ad set level. It optimizes delivery for incremental conversions using machine learning models that predict whether a conversion was caused by an ad, and reports on that basis. That is closer to the question finance asks, but it is a prediction, not an experiment in your account. Meta also warns that each attribution model counts conversions differently, so ad sets on different models shouldn't be compared side by side in the Campaign Overview table.
Weak tracking hides the gap rather than closing it. For people on iOS 14.5 and later, Meta says Aggregated Event Measurement recovers limited measurement of website conversions. Attributed counts can miss sales the ads drove and still credit sales they didn't. A different attribution window changes what gets counted, not what was caused; our attribution settings guide covers the windows. Blended MER doesn't depend on the pixel, but it mixes every channel together and can't isolate Meta, as our ROAS vs MER guide explains.
A CFO needs a causal number tied to money: incremental revenue per dollar, or cost per incremental customer, with a range around it and a decision rule agreed before the test. Those numbers slot into the CAC and LTV math in our unit economics guide. Four tools come up in this work, and each answers a different question.
- A/B test: which strategy costs less per result, on split audiences
- Conversion Lift: how many conversions Meta ads caused, against a randomized holdout of people
- Geo holdout: the same question answered by region, read in your own sales
- Marketing mix model: each channel's contribution over time, calibrated by the experiments
Meta A/B tests: comparing strategies on split audiences
An A/B test compares versions of a strategy that differ in one variable: creative, audience, placement or a whole campaign approach. Each version runs to a separate segment of your audience so nobody sees both; Meta's Marketing API documentation describes mutually exclusive audiences with no overlap between groups. Meta generally recommends building one from the Ads Manager toolbar, or you can use Experiments, which compares up to five existing campaigns. Meta advises against informal tests, such as switching ad sets on and off by hand, which can overlap audiences and give unreliable results.
Meta's best practices for a clean read: test one variable, starting from a measurable hypothesis. Define the KPI and confidence level before creating the test, and give both versions the same budget. Don't use the test audience in any other campaign running at the same time, since overlapping audiences can contaminate results. Run for at least seven days, within the 30-day maximum, and longer if your customers typically take more than a week to convert. Once a test is running, you can't change the variable it tests.
Know what the result means. Meta picks the winner on cost per result for the event you chose, simulating outcomes tens of thousands of times to attach a confidence percentage. Its lift documentation notes that A/B tests don't include randomized holdout groups. Meta does let you add a holdout to an A/B test and compare cost per incremental purchase; without one, the winner is the cheaper strategy per attributed result, not proof that either version is incremental. Use A/B tests to choose creative, audiences and settings, and lift to judge whether Meta spend pays for itself.
Conversion Lift: randomized holdouts and how access works
Conversion Lift is Meta's randomized experiment. Meta randomly splits the audience into a test group that can see your ads and a control group that can't, and its Business Help Center says people stay in their group for the whole test. Meta counts everyone in the test group, including people who never scrolled to an ad, which keeps the groups comparable. Meta's lift API lets one study hold several test cells, each with its own holdout, to compare approaches.
Conversions are matched to both groups from your pixel, Conversions API, app SDK or offline uploads, and Meta recommends the Conversions API because it depends less on the browser. Results come with intervals showing the range of most probable outcomes. Keep them separate from Ads Manager. Lift counts all conversions in both groups between the test dates and scales the holdout to the test group's size, while Ads Manager counts conversions attributed inside windows; for those reasons Meta advises against comparing the two.
Access runs on two tracks. Meta's Business Help Center says you can create a test in Experiments, or through your Meta account representative if you have one, with no charge for the test itself, though its campaigns may have budget requirements. As a guide, the account needs a campaign started in the past year with at least USD 5,000 in spend and 500 conversions, prorated for campaigns longer than 90 days. Self-serve tests also need the Conversions API sending an event with event match quality above 5, or a supported data source sending a low-funnel or app install event, a route Meta says isn't available to advertisers mainly targeting countries affected by the EU ePrivacy Directive. Through the Marketing API, Conversion Lift access is currently limited and arranged through a Meta representative, and Meta assigns representatives proactively, with no process to request one.
Geo holdouts when a lift study isn't available
A geo holdout swaps people for places. You change Meta spend in a set of test regions, switching it off for a holdout or adding it for a new campaign, keep comparable regions unchanged, and compare sales in your own data. Because the read comes from your order system rather than the pixel, it works when tracking is weak or events are restricted. The tradeoff is precision: Meta recommends people-based experiments where possible, GeoLift's documentation says, because they have higher statistical power.
GeoLift is an open-source R package published by Meta Open Source, and its site states that it is not a Meta product and is for research purposes only. Its documentation describes a Marketing Science research tool for cases where people-based measurement isn't feasible, naming healthcare advertisers as an example. It builds a synthetic control, a weighted mix of untreated regions that tracks your test regions before launch, and reads lift as the gap afterwards. The docs warn that a geo test without a prior power analysis has a high chance of missing real lift, and include power calculators for test length, investment and market selection.
Its best practices set the bar: daily data where possible, at least 25 pre-treatment periods across 20 or more geographic units, stable history of four to five times the test length, ideally 52 weeks, and no missing values. Run at least 15 days on daily data, or four to six weeks on weekly, covering at least one purchase cycle. Match test and control regions on the outcome you will measure, and hold other local and national marketing constant.
Health and wellness brands often end up here. Meta can assign websites, apps and other data sources to categories, health and wellness among them, whose data sharing restrictions can block specific mid- and lower-funnel events, or all events, in some or all regions. A lift study can't count purchases Meta never receives, so a geo test read against backend sales may be the only causal measure left; our health and wellness restrictions guide covers the restriction levels.
Marketing mix modeling with Robyn
A marketing mix model answers a standing question rather than a one-off one: how much each channel contributes to sales, allowing for lagged effects and diminishing returns, from aggregate data instead of user tracking. Robyn is Meta Marketing Science's experimental, open-source MMM package. It combines ridge regression, Nevergrad hyperparameter search and Prophet decomposition of trend and seasonality, and models adstock, the lag and decay of ad effects, and saturation, where each extra unit of exposure adds less response. It needs no personal or user-level data and doesn't depend on cookies or pixel data.
The data bar is high. Robyn's analyst guide asks for at least two years of weekly history and about one input variable per ten observations, so two years of weekly data supports only around ten variables. Spend has to vary, too: if a channel's spend stayed flat, the model struggles to tell what it did. And Robyn is code your analysts run and tune, not a report in Ads Manager.
Calibration is what makes it credible to finance. Robyn's guide strongly recommends calibrating with experiments, naming Meta Conversion Lift and GeoLift, entered as point estimates, and says to avoid studies whose confidence is not high enough. Its budget allocator then solves for the spend mix that maximizes response and shows which channels to raise or cut, though the guide says there's no guarantee its predictions will meet business expectations and asks you to validate outputs first. Treat a big recommended shift as the hypothesis for your next test.
Designing a test that can reach a decision
A test is only as useful as the decision it was built to make. Write down the question, the hypothesis, the primary KPI in backend terms, the confidence level and the decision rule, and have finance sign it. The rule names an action for each outcome: scale if the low end of the incremental ROAS range clears break-even, cut if the high end misses it, rerun bigger or longer if it straddles. Without one, every result gets relitigated after the fact.
Size the test from published guidance: Meta's seven-to-30-day window for A/B tests and GeoLift's power calculators for geo tests. Meta's Conversion Lift thresholds describe which accounts qualify, not how large a test must be, and a test too small to detect the lift you need is wasted spend. Then freeze it. Meta says editing during the learning phase resets learning, a running A/B test locks its variable, and a started lift study can't change its start time or treatment percentages or drop ad accounts and campaigns. Hold promotions and pricing steady, and skip the post-test conversion window, which Meta doesn't recommend for calculating results.
Contamination is the quiet failure. Ads reaching the holdout from outside the design, such as prospecting from another ad account in a holdout region, raise the control group's sales and make Meta look weaker than it is. Exclude holdout regions in every ad account and campaign that could reach them, not only the campaign under test.
- Question, hypothesis and primary KPI, measured in backend orders or revenue
- Decision rule and confidence level, signed by finance before launch
- Length and budget from Meta's guidance or a GeoLift power analysis
- Every ad account and campaign that could reach the holdout, with exclusions applied
- A change freeze on tested campaigns, and one owner for the read
Turning lift results into budget and scaling decisions
Translate the result into finance's numbers. Incremental ROAS is incremental revenue divided by the spend that produced it; cost per incremental conversion is that spend divided by incremental conversions. Judge both against your own targets rather than Ads Manager: break-even ROAS from contribution margin, and for subscription or lead businesses, the CAC your LTV supports, which our unit economics guide covers. Say a cell spent $150,000 and the study estimates $330,000 of incremental revenue: incremental ROAS is 2.2.
At a 40% contribution margin, break-even is 2.5, so that spend loses money on first orders even if Ads Manager shows a 4.0 ROAS, and only repeat purchases can justify it. Read the range, not the point estimate: if the low end clears your target, scale; if the high end misses it, cut or rebuild. A lift result also measures the average return on the spend tested, and saturation means the next dollar returns less, so test again before a big step up.
Then ramp in steps. Meta says frequent budget changes can push ad sets back into the learning phase, so raise spend gradually and check backend sales between steps; our budget increase rules guide covers pacing and our learning phase guide covers resets. Retest after major changes, such as a new offer, market or campaign structure, and feed each result into the MMM.
Running incrementality tests across many ad accounts
At $100k a month and up, spend often runs across several ad accounts, split by market, product, client or team, and any of them can reach the people or regions under test. So the test has to cover the advertiser, not one account. In Meta's lift API, a single test cell can list several ad accounts, and Meta runs and measures all ads under the cell's active ad entities during the study. List every account that can reach the test audience.
Separate accounts are not a split. Putting the test strategy in one account and the control in another randomizes nothing, since the same people can see both; divide audiences with Experiments, the API or geography. For geo holdouts, apply exclusions in every ad set of every account, untick "Reach more people likely to respond to your ads", which can be checked by default on city and region targeting, and copy exclusions into any account opened mid-test. Meta notes that someone living in an included location may still see ads while visiting an excluded one, and that location accuracy isn't guaranteed, so expect some leakage.
Keep one test calendar across accounts. Experiments won't let an A/B test use a campaign that is already in another test, but overlapping audiences across accounts are yours to catch. Give one owner approval over every test, and export spend by account, region and day next to backend sales, the input a geo read needs.
When a positive read justifies more spend, Power Ads supplies the account capacity: agency ad accounts created in its own business portfolio and shared into your Business Manager through partner access, on its corporate cards, with unlimited accounts and free replacements, for 4% deducted from each top-up and no setup, onboarding, monthly or per-account fee. It supplies accounts only; test design, lift studies, measurement and tracking stay with your team, and every account it shares belongs inside your test design.
Key takeaways
- Standard attribution credits conversions inside a window and incremental attribution estimates causation with Meta's own model; only a comparison with people or regions that couldn't see your ads measures incremental revenue.
- Meta A/B tests split audiences with no overlap and pick the cheaper strategy on one variable, but without a holdout they don't prove incrementality.
- Conversion Lift randomly splits people into test and control groups; it is self-serve in Experiments for eligible accounts, while Marketing API access is limited and arranged through a Meta representative.
- Geo holdouts read in backend sales work when people-based lift isn't feasible, as with restricted health datasets; GeoLift is open-source, research-only and not a Meta product.
- Robyn, Meta Marketing Science's experimental open-source MMM, needs about two years of weekly data and should be calibrated with lift or geo results.
- Write the hypothesis and decision rule first, freeze the test, exclude holdouts in every ad account, and scale on incremental ROAS against your MER and break-even targets.
FAQ
How do I prove to my CFO that Meta ads drive revenue when tracking is broken?
Show a causal number from a test agreed in advance. Run Conversion Lift if your account qualifies, or a geo holdout read in your own sales data, then report incremental revenue and incremental ROAS, with the range around them, against the decision rule finance signed. Neither depends on every sale being attributed. For a standing cross-channel view, calibrate a marketing mix model such as Robyn with those results.
What is the difference between Meta ads attribution and incrementality?
Standard attribution credits a conversion to an ad if it happened within a set number of days after someone viewed or clicked it, so it can include sales that would have happened anyway. Incrementality counts only the conversions the ads caused, measured against a group that couldn't see them. Meta's incremental attribution model predicts which conversions ads caused, which helps optimization, but it is a model, not a test. Because the methods count conversions differently, Meta advises against comparing Conversion Lift results with Ads Manager reporting.
Can I run a Conversion Lift test without a Meta representative?
Yes, if your account qualifies. Meta's Business Help Center says eligible accounts can create one in Experiments; its guide is a campaign started in the past year with at least USD 5,000 in spend and 500 conversions, plus a Conversions API or low-funnel signal requirement for self-serve tests. Marketing API access to Conversion Lift is currently limited and goes through a Meta representative, and Meta says there is no process for advertisers to request one.
Is GeoLift an official Meta product?
No. GeoLift is an open-source R package published by Meta Open Source, and its site says it is not a Meta product and is for research purposes only. Its documentation positions it for cases where people-based lift isn't feasible, such as healthcare advertisers, and notes that Meta recommends people-based experiments where possible because they have higher statistical power.
How long should a Meta incrementality test run?
It depends on the method. For A/B tests, Meta recommends at least seven days, allows up to 30, and suggests running longer if customers take more than a week to convert, though an A/B test only measures incrementality with a holdout. GeoLift's best practices call for at least 15 days on daily data, or four to six weeks on weekly data, covering at least one purchase cycle, with a power analysis to confirm length and budget. For lift studies, Robyn's guide suggests longer durations to capture long-tail conversions and reduce bias.
Does Power Ads run lift tests or set up tracking?
No. We supply agency ad accounts, created in our business portfolio and shared into your Business Manager through partner access, on our corporate cards, for 4% deducted from each top-up. Test design, lift studies, geo holdouts, MMM and tracking stay with your team or your measurement partner, and every account we share should sit inside your test design.
Sources
- facebook.com/business/help/221353413010930
- facebook.com/business/help/552097218528551
- facebook.com/business/help/365383970855344
- facebook.com/business/help/1471026819588421
- developers.facebook.com/docs/marketing-api/guides/lift-studies
- facebook.com/business/help/460276478298895
- facebook.com/business/help/2366718460372682
- facebook.com/business/help/458681590974355
- facebook.com/business/help/721422165168355
- facebook.com/business/help/1738164643098669
- facebook.com/business/help/1159714227408868
- facebook.com/business/help/3506622486044209
- facebook.com/business/help/290009911394576
- facebook.com/business/help/166313650471318
- facebook.com/business/help/286876155019610
- developers.facebook.com/docs/marketing-api/guides/split-testing
- facebookincubator.github.io/GeoLift
- facebookincubator.github.io/GeoLift/docs/intro
- facebookincubator.github.io/GeoLift/docs/Methodology
- facebookincubator.github.io/GeoLift/docs/Best%20Practices/BestPractices
- facebookexperimental.github.io/Robyn
- facebookexperimental.github.io/Robyn/docs/analysts-guide-to-MMM
- facebook.com/business/help/511197658391698
- facebook.com/business/help/1402913027039332
- facebook.com/business/help/365561350785642
- facebook.com/business/help/112167992830700
