What marketing incrementality measures
Marketing incrementality is the causal change in an outcome produced by a marketing intervention. The outcome might be revenue, orders, sign-ups, conversions, store visits, or another business KPI. The important word is causal: an incremental conversion is one that would not have happened in the absence of the campaign.
This distinction matters because customers who see or click ads are often already more likely to buy. Brand demand, seasonality, promotions, distribution changes, competitor activity, and macroeconomic conditions can all move outcomes at the same time as media spend. Counting exposed or attributed conversions does not separate these forces.
Observed performance is known. Incrementality estimates the missing world: what would performance have been if the campaign had not run?
Incrementality versus attribution
Attribution assigns credit across recorded interactions. Incrementality estimates the difference caused by an intervention. They are useful for different jobs and should not be treated as interchangeable.
| ATTRIBUTION | INCREMENTALITY | |
|---|---|---|
| Primary question | Which touchpoint receives credit? | Did marketing cause additional outcomes? |
| Comparison | Rules or model across observed paths | Observed outcome versus counterfactual |
| Typical output | Attributed conversions and ROAS | Incremental lift, iROAS, iCAC and uncertainty |
| Main risk | Crediting demand that already existed | Weak controls, spillover or an underpowered test |
Attribution can still support operational decisions such as creative diagnostics or journey analysis. Incrementality is better suited to budget allocation, channel validation, and questions where the cost of over-crediting marketing is material.
Why the counterfactual is the hard part
The test market cannot be observed in both treated and untreated states at the same time. Measurement therefore needs a credible substitute for the missing untreated outcome. This is the counterfactual.
In OpenLift, the counterfactual is predicted from matched control markets. A model learns the pre-treatment relationship between the test geography and controls, then projects that relationship into the treatment period. The difference between the observed test-market outcome and posterior counterfactual is estimated lift.
Common experiment designs
Randomised holdouts
Users, households, or geographies are randomly assigned to treatment and control. Randomisation is the strongest defence against systematic differences, but it may be unavailable because of platform constraints, privacy, channel mechanics, or operational risk.
Geo-lift tests
Marketing changes in selected regions while comparable regions remain untreated. Geo tests work well when user-level randomisation is impossible and spend can be controlled geographically. Their validity depends on market comparability, stable measurement, limited spillover, and sufficient power.
Quasi-experimental designs
Difference-in-differences, synthetic controls, and interrupted time series use observational variation to estimate impact. These methods require explicit assumptions and careful diagnostics; the model does not make a weak design causal by itself.
From lift to a business decision
A useful result is more than a single percentage. OpenLift combines the posterior lift distribution with match quality and data-quality diagnostics, then translates the effect into business economics when spend and assumptions are supplied.
- Probability of positive lift expresses how much posterior mass sits above zero.
- Credible interval shows the plausible range of effects under the model.
- Incremental ROAS divides incremental revenue by spend.
- Incremental CAC divides spend by incremental customers.
- Incremental profit incorporates margin rather than treating revenue as value.
The decision should respond to both expected impact and uncertainty. A positive mean with a wide interval may justify a better-powered retest, not immediate scale.
What incrementality cannot guarantee
No method produces perfect attribution or guaranteed true ROAS. Geo experiments can be weakened by poor control matches, short test windows, sparse outcomes, volatile revenue, overlapping campaigns, tracking changes, and treatment spillover. Economic outputs are only as reliable as the margin, lifetime value, and cost inputs supplied.
The defensible interpretation is evidence-based: estimate whether marketing likely caused additional growth, make the uncertainty visible, and choose an action proportionate to the strength of the experiment.