Attributed revenue answers which touchpoint received credit under a chosen rule. Incrementality asks a harder question: how much changed because the marketing happened? The answer depends on a credible estimate of what would have happened otherwise. Experiments can create a comparison group; geographic designs and statistical models can help when randomization is difficult. Each method has assumptions, and each can be misread as more certain than it is. This guide separates lift from attribution and shows how to connect causal evidence to the next budget decision.
🧭 Jump to a term
Define the causal question and outcome: Incrementality · Incremental sales · Counterfactual
Choose a comparison design: Holdout experiment · Geo-lift test · Matched-market test · Synthetic control · Selection bias · Confounding variable
Measure different kinds of lift: Conversion lift · Brand lift · Incremental return on ad spend (iROAS)
Model persistence and the next investment: Marketing mix modeling (MMM) · Adstock · Media saturation curve · Marginal return on investment (marginal ROI)
⭐ Know these first
Start with Incrementality, Incremental return on ad spend (iROAS), Holdout experiment, Geo-lift test, Marketing mix modeling (MMM). Then follow the grouped learning order below.
📎 How to read this page
What it means gives the precise meaning. Operator translation gives the version you might hear in a real ecommerce meeting. In real life shows an illustrative example. Watch out and the confusion boxes show where a familiar term can mislead.
📈 Read the relationships first
These combinations are diagnostic hypotheses, not proof of causality. Compare the same period and scope, then investigate the mechanism.
Attribution ↑ + randomized lift ↓
Usually means: more sales receive channel credit than a test estimates were caused by the channel. Check next: eligibility, overlap, organic demand and holdout design.
Media spend ↑ + marginal return ↓
Usually means: the next budget increment may be less productive than earlier spend. Check next: saturation, audience overlap, channel constraints and uncertainty.
Control contamination ↑ + estimated lift ↓
Usually means: exposed and unexposed groups may no longer represent distinct conditions. Check next: cross-device exposure, geo spillover and suppression coverage.
Define the causal question and outcome
01 · 🟢 Core
Incrementality = Causal impact beyond the counterfactual
🧠 What it means
The causal effect of a marketing activity: the outcomes that occurred because of it compared with a credible estimate of what would have happened without it.
💬 OPERATOR TRANSLATION
“A dashboard can count a sale; a test asks whether the ad made the difference.”
🛍️ In real life
A randomized holdout estimates how many purchases would not have happened without the campaign.
Check next
Use lift evidence to assess whether spend created additional outcomes.
🔗 Related: Incremental sales · Counterfactual · ↑ all terms
02 · 🔵 Operations
Incremental sales = Sales caused beyond the baseline
🧠 What it means
Sales above the estimated counterfactual that are caused by a defined marketing intervention.
💬 OPERATOR TRANSLATION
“Sales that were going to happen anyway are not a marketing miracle.”
🛍️ In real life
A test estimates 800 purchases in the exposed group versus 720 comparable purchases without the campaign: 80 incremental purchases.
Check next
Translate incremental units or revenue into contribution after discounts and costs.
🔗 Related: Incrementality · Counterfactual · ↑ all terms
03 · 🔵 Operations
Counterfactual = Outcome under the alternative condition
🧠 What it means
The unobserved outcome that would have occurred for the same unit or population under the alternative condition.
💬 OPERATOR TRANSLATION
“The missing version of reality is the whole measurement problem.”
🛍️ In real life
The campaign’s outcome is compared with the estimated result had eligible customers not received the ads.
Check next
State how the counterfactual was built and where it may fail.
🔗 Related: Incrementality · Incremental sales · ↑ all terms
Choose a comparison design
04 · 🟢 Core
Holdout experiment = Treatment compared with a withheld group
🧠 What it means
A design that withholds a treatment from a randomly assigned or otherwise comparable group to estimate the treatment’s effect.
💬 OPERATOR TRANSLATION
“A control group needs a plan before the result arrives.”
🛍️ In real life
A random share of eligible customers is suppressed from an email campaign and compared with those sent it.
Check next
Predefine the population, assignment, outcome, duration and analysis.
🔗 Related: Geo-lift test · Matched-market test · Synthetic control · ↑ all terms
05 · 🟢 Core
Geo-lift test = Compare treated and control geographies
🧠 What it means
A geographic experiment that compares outcomes in treated areas with outcomes in control areas after a marketing change.
💬 OPERATOR TRANSLATION
“A state line is not a magic randomizer.”
🛍️ In real life
A retailer increases streaming ads in selected regions and compares sales changes with matched control regions.
Check next
Check pre-period fit, leakage, market count and other local changes.
🔗 Related: Holdout experiment · Matched-market test · Synthetic control · ↑ all terms
06 · 🔵 Operations
Matched-market test = Compare matched markets
🧠 What it means
An experiment that pairs comparable markets and assigns treatment to one market in each pair.
💬 OPERATOR TRANSLATION
““These towns look alike” should be a measured claim.”
🛍️ In real life
Pairs of cities with similar prior sales are formed; one of each pair receives a new campaign.
Check next
Document matching variables and assignment method, then report uncertainty.
🔗 Related: Holdout experiment · Geo-lift test · Synthetic control · ↑ all terms
07 · 🔵 Operations
Synthetic control = Weighted estimate of a treated unit’s counterfactual
🧠 What it means
A comparative method that constructs a weighted combination of untreated units to approximate a treated unit’s counterfactual.
💬 OPERATOR TRANSLATION
“A synthetic twin still needs a convincing pre-treatment resemblance.”
🛍️ In real life
A treated region is compared with a weighted combination of other regions that matched its pre-campaign trend.
Check next
Review pre-period fit, donor pool, placebo checks and uncertainty.
🔗 Related: Holdout experiment · Geo-lift test · Matched-market test · ↑ all terms
08 · 🔵 Operations
Selection bias = Distortion from how groups are selected
🧠 What it means
Systematic error that occurs when inclusion or treatment assignment is related to factors that also affect the measured outcome.
💬 OPERATOR TRANSLATION
“The easiest people to reach may also be the easiest people to convert.”
🛍️ In real life
High-intent shoppers are more likely to see retargeting ads and also more likely to buy without them.
Check next
Use randomization or a defensible adjustment and test remaining imbalance.
🔗 Related: Holdout experiment · Geo-lift test · Matched-market test · ↑ all terms
09 · 🔵 Operations
Confounding variable = Third factor affecting treatment and outcome
🧠 What it means
A factor associated with both the treatment and outcome that can distort an estimated relationship.
💬 OPERATOR TRANSLATION
“If the ad and the sale rose together, check what else moved.”
🛍️ In real life
A holiday promotion coincides with an ad increase and independently raises demand.
Check next
Log major concurrent changes and choose a design that can distinguish their effects.
🔗 Related: Holdout experiment · Geo-lift test · Matched-market test · ↑ all terms
Measure different kinds of lift
10 · 🔵 Operations
Conversion lift = Causal change in conversions
🧠 What it means
The change in a defined conversion outcome caused by an advertising exposure or intervention, estimated against a comparison group.
💬 OPERATOR TRANSLATION
“Ten percent lift sounds different when the base is two in a hundred.”
🛍️ In real life
Conversion rises from 2.0% in control to 2.2% in treatment: +0.2 percentage points, or +10% relative.
Check next
Use the absolute effect for volume planning and uncertainty for decisions.
🔗 Related: Brand lift · Incremental return on ad spend (iROAS) · ↑ all terms
11 · 🔵 Operations
Brand lift = Change in measured brand outcome
🧠 What it means
A measured change in brand outcomes such as awareness or consideration associated with a campaign, typically compared with an unexposed group.
💬 OPERATOR TRANSLATION
“A brand metric can move before revenue, but it is still not revenue.”
🛍️ In real life
An exposed survey group reports higher aided awareness than a matched unexposed group.
Check next
Pair brand metrics with the decision horizon and downstream evidence.
🔗 Related: Conversion lift · Incremental return on ad spend (iROAS) · ↑ all terms
12 · 🟢 Core
Incremental return on ad spend (iROAS) = Incremental revenue per ad dollar
🧠 What it means
Incremental revenue attributed causally to advertising divided by the advertising spend for the defined activity and period.
💬 OPERATOR TRANSLATION
“Four dollars of extra sales per ad dollar is not automatically four dollars of profit.”
🛍️ In real life
An experiment estimates $120,000 incremental revenue from $30,000 media spend: iROAS is 4.0 on that basis.
Check next
Compare iROAS with contribution margin and the business hurdle.
🔗 Related: Conversion lift · Brand lift · ↑ all terms
Model persistence and the next investment
13 · 🟢 Core
Marketing mix modeling (MMM) = Model of marketing and business outcomes
🧠 What it means
A statistical approach that estimates how marketing and other factors relate to aggregate business outcomes over time, often to evaluate contributions and scenarios.
💬 OPERATOR TRANSLATION
“The model can connect the dots; it cannot invent variation that never happened.”
🛍️ In real life
A model uses weekly sales, media spend, pricing, promotions and seasonality to estimate channel contribution.
Check next
Review holdout validation, uncertainty, granularity and decision use before reallocating.
🔗 Related: Adstock · Media saturation curve · Marginal return on investment (marginal ROI) · ↑ all terms
14 · 🔵 Operations
Adstock = Modeled advertising carryover
🧠 What it means
A model representation of advertising effects that persist or decay over time after exposure.
💬 OPERATOR TRANSLATION
“An ad can linger; the model decides how long the echo lasts.”
🛍️ In real life
A campaign’s estimated effect carries into later weeks with decreasing weight.
Check next
Inspect carryover assumptions and sensitivity across plausible decay patterns.
🔗 Related: Marketing mix modeling (MMM) · Media saturation curve · Marginal return on investment (marginal ROI) · ↑ all terms
15 · 🔵 Operations
Media saturation curve = Response as spend increases
🧠 What it means
A modeled relationship in which the incremental outcome from additional media investment changes as spend or exposure increases.
💬 OPERATOR TRANSLATION
““We can scale” is a curve question, not a mood.”
🛍️ In real life
A model estimates strong early gains from a channel, followed by diminishing incremental returns.
Check next
Use uncertainty ranges and controlled tests near budget decisions.
🔗 Related: Marketing mix modeling (MMM) · Adstock · Marginal return on investment (marginal ROI) · ↑ all terms
16 · 🔵 Operations
Marginal return on investment (marginal ROI) = Return from the next investment
🧠 What it means
The additional return expected from the next unit of investment, measured against its incremental cost.
💬 OPERATOR TRANSLATION
“Average performance does not tell you what the next dollar will do.”
🛍️ In real life
The next $10,000 of media is expected to add $14,000 contribution: incremental return is $4,000, or 40% of cost.
Check next
Compare the next-dollar estimate across feasible alternatives and uncertainty.
🔗 Related: Marketing mix modeling (MMM) · Adstock · Media saturation curve · ↑ all terms
🔀 Attribution vs incrementality
Attribution assigns credit under a rule; incrementality estimates causal change against a counterfactual.
🔀 iROAS vs platform ROAS
iROAS uses incremental revenue; platform ROAS typically uses attributed revenue under the platform’s measurement rules.
🔀 Percentage-point lift vs relative lift
A move from 2.0% to 2.2% is +0.2 percentage points and +10% relative. State which one you mean.
🔀 MMM vs randomized experiment
MMM models aggregate historical relationships and scenarios; an experiment compares outcomes under assigned conditions. They answer related but different questions.
🤔 Still confused?
Follow this thread: Incrementality → Holdout experiment → Conversion lift → Marketing mix modeling (MMM). That sequence moves from the basic object or relationship to the decisions and checks it supports.
Sources and scope
Primary documentation checked on 27 September 2026. Platform features and eligibility can change; examples and cartoon situations are illustrative.
