A programmatic report can show that a conversion happened after an impression. It cannot, by itself, show that the impression caused the conversion. That gap matters most in retargeting, brand search, loyalty-heavy retail and any campaign aimed at people already likely to buy.
Incrementality measurement closes the gap by estimating the counterfactual: what would have happened without the advertising? The right design may be a randomized user holdout, a geo experiment, platform Conversion Lift, ghost ads or a calibrated marketing mix model. The method depends on how media is bought, how outcomes are observed and how much signal is available.
Quick answer: To prove programmatic caused incremental conversions, create a credible counterfactual with a randomized holdout or matched geo test. Keep treatment and control comparable, prevent media overlap, predefine the conversion window and analyze absolute lift, relative lift, confidence intervals, incremental CPA and incremental ROAS. Attribution alone cannot establish causality.
How do the main programmatic measurement methods compare?
Method | What it estimates | Best use | Main risk
|
|---|---|---|---|
Attribution | Credit among observed touchpoints | Daily reporting and journey diagnostics | Confuses correlation with causation |
User holdout / Conversion Lift | Difference between eligible exposed and withheld users | Platform or campaign-level causal lift | Identity loss, contamination and insufficient conversions |
Geo-lift | Incremental outcome across matched regions | Cross-channel, offline and privacy-durable measurement | Geo imbalance, spillover and external shocks |
Ghost ads / ghost bids | Lift among users who would have won or been served an ad | Auction-aware display and retargeting tests | Requires platform or log-level implementation |
MMM | Channel contribution and response curves over time | Portfolio planning and channels that cannot be randomized | Model dependence, confounding and wide uncertainty |
No method replaces the others. Attribution is fast and granular. Experiments are stronger for causal questions. MMM covers more channels and longer planning horizons. A mature measurement system connects them instead of forcing one model to answer every question.
What is incrementality, and why is it different from attribution?
Attribution asks, “Which touchpoint gets credit?” Incrementality asks, “How many outcomes would not have happened without the media?” Those are different questions.
Imagine 1,000 purchasers were exposed to a retargeting campaign and 200 converted. An attribution report may credit many of those 200 conversions to programmatic. But if 180 would have converted without the campaign, the campaign caused only 20 incremental conversions. The attributed count is 200; the incremental count is 20.
This is why last-click, view-through and data-driven attribution should not be treated as causal proof. They observe paths among people who received media. They do not directly observe the same people's behavior in a world where the media was withheld. Even a clean DV360 reporting metrics setup still needs a control condition to answer the causal question.
Incrementality also changes the economic metrics. Cost per attributed acquisition divides spend by credited conversions. Incremental CPA divides spend by incremental conversions. Attributed ROAS uses credited revenue. Incremental ROAS uses the additional conversion value caused by media. These figures can diverge sharply, especially when targeting high-intent audiences.
How do you prove programmatic caused incremental conversions?
The strongest practical proof is a controlled experiment with a valid counterfactual. Define the eligible population before the campaign. Randomly assign comparable users, households or regions to treatment and control. Serve normal programmatic media to treatment. Withhold the tested media from control. Measure the same conversion event over the same window. The difference in outcomes is the estimated lift.
For equal-sized groups, the core calculation is simple:
Absolute lift = treatment conversions - control conversions
Relative lift = (treatment conversion rate - control conversion rate) / control conversion rate
Incremental CPA = test spend / incremental conversions
Incremental ROAS = incremental conversion value / test spend
If groups differ in size or exposure opportunity, compare rates or scale the control outcome to the treatment population. Report an interval around the estimate, not only a point value. A lift of 8% with a wide interval crossing zero is inconclusive, not proof of no effect and not permission to present 8% as certain.
Google's official Conversion Lift documentation follows this treatment-versus-control logic. Its Display & Video 360 guidance says control users receive the next ad in the auction, while treatment users may receive the tested ad. That design isolates direct impact more credibly than comparing exposed users with naturally unexposed users.
Which experiment should you choose?
Use a randomized user holdout when identity and platform support are strong
User-level holdouts offer the cleanest randomization and usually the best statistical efficiency. They fit large digital campaigns with reliable conversion events and a platform that can suppress ads for the control group. Google's user-based Conversion Lift is one example.
Watch for identity gaps, logged-out traffic, cookie churn and cross-device conversions. Control users may also see similar ads from another campaign or DSP. That contamination usually pushes measured lift toward zero.
Use geo-lift when outcomes or media cannot be separated by user
Geo experiments assign regions to treatment and control, then compare the change in sales, leads, visits or another KPI. They work well for CTV, DOOH, retail sales, offline conversions and cross-channel campaigns. They are also useful when privacy restrictions make user-level matching incomplete.
Do not pair geos by population alone. Match them on pre-period outcome level, trend, volatility, seasonality and media conditions. Google's Meridian GeoX documentation recommends stratified sampling and evaluates candidates using out-of-sample fit, A/A robustness, minimum detectable effect and budget. Prevent spillover from commuting, shipping, overlapping broadcast markets or national media.
Use ghost ads or ghost bids when the auction creates selection bias
In a normal ad experiment, “unexposed” users may differ from exposed users because the DSP did not bid, lost the auction or predicted low value. Ghost-ad methods preserve the auction decision. The control group records the impression opportunity that would have occurred, but withholds the tested ad. This compares exposed users with control users who were genuinely eligible for the same auction outcome.
The method can improve precision and reduce the cost of public-service-ad controls, but it needs platform cooperation or detailed auction logs. The 2017 Journal of Marketing Research paper that introduced ghost ads describes both observed and predicted ghost-ad designs. Do not label an ordinary exposed-versus-unexposed report “ghost bidding.”
Use MMM when the question is portfolio-level
Marketing mix modeling uses variation over time, and often geography, to estimate channel contribution, response curves and ROI. It is useful for budget planning across programmatic, search, social, TV, promotions and other business drivers. Google's open-source Meridian is designed for causal inference and can use experiment results as priors.
MMM is not a shortcut around experimentation. Results depend on data quality, controls, model assumptions and available variation. Calibrate the model with strong lift tests where possible. Use MMM for portfolio decisions and experiments for the campaign, audience or tactic questions the model cannot resolve cleanly.
What experiment-selection framework should a team use?
Define the decision. Are you deciding whether to fund programmatic, which DSP to use, whether retargeting adds value or how to allocate the full media budget?
Choose the unit you can withhold. Use users or households when identity and suppression are reliable. Use geos when outcomes are aggregated, offline or cross-channel.
Check auction access. If auction selection is central and the platform supports it, use ghost ads or a platform-native randomized lift product.
Check signal. Estimate baseline conversions, expected lift, variance and feasible duration. If the test is badly underpowered, broaden the conversion event, increase the eligible population or move up to geo or MMM.
Match method to scope. Use experiments for a bounded causal question. Use MMM for recurring portfolio allocation. Use attribution for operational diagnostics.
A practical shortcut: if you can randomize cleanly, randomize. If you cannot randomize users but can vary media geographically, run geo-lift. If neither is feasible, use a carefully specified quasi-experiment or MMM and be explicit that the evidence is more assumption-dependent.
How should you design a valid programmatic lift test?
Write the analysis plan before launch. State the primary KPI, experiment unit, allocation ratio, eligibility rules, expected conversion lag, test duration, attribution window, exclusion rules and stopping rule. Register one primary decision metric. Secondary breakdowns are useful, but repeated slicing increases the chance of finding noise that looks significant.
Keep every element except the tested media as stable as possible. Treatment and control should share pricing, product availability, promotions, landing pages and CRM policy. Coordinate search, social, email and affiliate teams so control users or regions are not compensated with extra activity.
Instrumentation must also agree. Reconcile Floodlight, CRM and analytics definitions before the test. The article on DV360 traffic in GA4 explains why source labels can differ across systems. Use a stable event contract, deduplicate conversions and keep timestamps in a consistent time zone. Build a monitoring view with the DV360 Looker Studio guide, but freeze the test logic before looking at lift.
For DV360 Conversion Lift, Google's current help page lists a 7-day minimum and 56-day maximum study duration and says campaigns should generate enough conversion volume for significance. It also publishes platform-specific feasibility and budget guidance. Treat those values as product eligibility rules, not a substitute for a power calculation based on your own baseline rate and minimum detectable effect.
How much sample size and budget do you need?
There is no universal conversion threshold. Required sample depends on the baseline conversion rate, outcome variance, minimum lift worth detecting, control share, desired significance level and target power. Smaller true effects need larger samples. Rare conversions need more eligible users. Noisy geo outcomes need more regions, more pre-period data or a larger spend change.
Start with the business threshold. If a 3% lift would not change the budget decision, do not power the test to detect 1%. Set the minimum detectable effect to the smallest result that would change action. Then estimate required observations using historical event rates. For geo tests, run placebo or A/A simulations across candidate splits. Reject designs that frequently “detect” effects before treatment begins.
Budget is part of design, not an afterthought. A treatment group must receive enough incremental spend or reach to create a detectable contrast. A control group that is too small produces unstable estimates. A control group that is too large creates unnecessary opportunity cost. Power the test before negotiating the allocation.
Which interpretation errors make lift studies misleading?
Calling inconclusive results “no lift.” A wide confidence interval means the study could not distinguish the effect from noise.
Reporting relative lift without the base. A 20% lift can be commercially small when the baseline conversion rate is tiny.
Stopping when results look good. Repeated peeking inflates false positives unless the design uses sequential methods.
Ignoring contamination. Other campaigns, CRM messages or cross-border exposure can reach the control group.
Changing optimization mid-test. Major bid, budget, audience or creative changes alter the treatment being estimated.
Using the wrong conversion lag. A short window misses delayed outcomes; an excessive window adds unrelated behavior.
Comparing incompatible estimands. A test of a 20% spend reduction estimates marginal impact, while an MMM zero-spend counterfactual estimates total channel contribution.
Complete the measurement preflight before launch: conversion mapping, audience overlap, frequency controls, brand safety and reporting permissions. The DV360 audit checklist is a useful operational companion because a broken setup cannot be repaired statistically after the test.
How should attribution, experiments and MMM work together?
Use attribution to diagnose where conversions appear and how paths change. Use randomized or geo experiments to estimate causal lift for high-value decisions. Use MMM to connect channels, account for broader business drivers and plan budgets over time.
The feedback loop matters. Experiment results can calibrate MMM priors. MMM can identify channels with high uncertainty or material spend that deserve the next test. Attribution can reveal operational issues and candidate segments, but it should not override a well-designed causal result merely because its credited conversion count is larger.
In reporting, show attributed conversions and incremental conversions as different measures. Add absolute lift, relative lift, uncertainty, incremental CPA and incremental ROAS. Record the tested population, dates, media scope and counterfactual. A decision-maker should be able to tell exactly what the estimate does and does not generalize to.
AdGeeks can map the counterfactual, select the experiment unit, calculate power, align DV360 and CRM data, and define the decision rules before spend begins.
Frequently asked questions
Can attribution prove incrementality?
No. Attribution assigns credit among observed touchpoints. Incrementality requires a credible estimate of what would have happened without the advertising, usually from a randomized holdout, geo experiment or carefully specified causal model.
What is a good incrementality lift?
A good lift is one that clears the business decision threshold after accounting for uncertainty and cost. Evaluate absolute conversions, incremental CPA and incremental ROAS, not a relative lift percentage in isolation.
What is control-group contamination?
Contamination occurs when control users or regions receive the tested advertising, a close substitute or coordinated marketing activity. It reduces treatment contrast and often biases measured lift toward zero.
When is geo-lift better than a user holdout?
Geo-lift is often better when conversions are offline, media spans channels, identity matching is weak or user-level suppression is unavailable. It needs comparable regions, enough pre-period data and controls for spillover.
Can MMM replace incrementality experiments?
No. MMM is valuable for portfolio planning and hard-to-randomize channels, but it depends on assumptions and observed variation. Experiments provide stronger calibration points and can test narrower campaign decisions directly.
Sources
Google Ads Help, About Conversion Lift, accessed October 20, 2026.
Google Ads Help, Comparing lift types, accessed October 20, 2026.
Display & Video 360 Help, Set up Conversion Lift measurement, accessed October 20, 2026.
Google for Developers, An introduction to Meridian, accessed October 20, 2026.
Google for Developers, Meridian GeoX design methodology, accessed October 20, 2026.
Johnson, Lewis and Nubbemeyer, “Ghost Ads: Improving the Economics of Measuring Online Ad Effectiveness”, Journal of Marketing Research, December 2017.
Brodersen et al., “Inferring causal impact using Bayesian structural time-series models”, Annals of Applied Statistics, 2015.
