Skip to content

Campaign Intelligence · Deep dive

Measure incrementality, not vanity metrics

Measure incrementality beyond last-click: geo and holdout tests for causality, attention beyond viewability and MMM calibrated on experiments. See the method.

Programmatic e Data — molti segnali, un solo piano

CAMPAIGN INTELLIGENCE · INCREMENTALITY

In · Raw signals: impressions, clicks, last-click conversions

Out · Incremental ROI you can defend

01

Counterfactual design (geo holdout / PSA-ghost control)

Builds the counterfactual by holding a control group out of exposure — geo holdout with matched markets or user-level PSA/ghost ads — so a world-without-the-campaign is observable and comparable, not merely assumed.

02

Attention layer (viewability baseline + attentive seconds)

Adds an attention layer measured in attentive seconds on top of the MRC viewability baseline, separating inventory that makes the message actually noticed from inventory that is technically visible but ignored.

03

Incrementality lift vs last-click credit

Computes the incremental lift — the conversions that would not have happened without the campaign — and contrasts it with the inflated last-click credit that simply intercepts demand already in motion.

04

MMM calibrated on experiment priors (Bayesian)

Feeds the experiment results into the marketing mix model as Bayesian priors, so the model corrects premature saturation and realigns strategic portfolio allocation, including offline and brand.

05

Triangulation: experiment + MMM + attribution

Triangulates experiment, MMM and attribution to cover each other's blind spots: the experiment anchors causality, the MMM gives the portfolio view, and attribution stays the granular diagnostic for creative and tactic.

06

One reading surface, one decision

Collapses everything into a single reading surface that yields one budget decision defensible to the CFO, reporting lift, confidence interval and statistical power instead of dashboards that contradict each other.

Signals · Symptoms

When the numbers climb but the business doesn't

If the reporting glows and revenue doesn't, you are usually measuring correlation, not cause. These are the tell-tale symptoms.

  • Last-click takes all the credit: lower-funnel channels look hyper-efficient because they intercept demand that would have converted anyway, proving no net-new value at all.
  • You pause a channel and sales hold: the clearest sign that spend was redundant, masked by a vanity metric crediting conversions that were never incremental.
  • High viewability, low recall: a technically visible ad is not a noticed ad; without an attention layer, viewability measures the opportunity to be seen, not being seen.
  • Every platform claims the same conversions: walled gardens report metrics designed to maximize spend inside their own walls, so the sum of credited results exceeds actual revenue.
  • Dashboards don't reconcile: attribution, brand tracking and models live in separate sheets and nobody reconciles the contradictions, so budget decisions stay opinions.

For CMOs, heads of growth and multinational media leaders who must defend budget allocation to the CFO with causal proof, not charts that go up.

Principles · Foundations

The right question: what would have happened anyway

Measuring incrementality means estimating the counterfactual: the difference between the world with the campaign and the world without it. Everything else is bookkeeping of the past.

Attribution ≠ causality

Last-click records the final measurable touch before conversion, but never estimates what would have happened without the ad. It records a path, not an effect. Useful for diagnostics, useless for deciding how much to spend.

Incrementality is net-new value

Incrementality measures the conversions that would not have occurred without the campaign. It is the only metric that answers the board's question: did every euro generate sales we would not otherwise have had?

The counterfactual is built, not assumed

Without a control group there is no incrementality, only conjecture. Geo holdouts, PSA/ghost ads and randomized splits exist to create an observable, comparable world-without-the-campaign.

Vanity metrics confuse volume with merit

Impressions, clicks and reach are inputs, not outcomes. They measure exposure, not behavior change. They turn dangerous the moment they enter bonus targets or scaling decisions.

No single method is complete

Experiments, models and attribution each have different blind spots. A reliable answer comes from triangulation, not from faith in one source.

Method · Experiment

Designing a test that survives internal cross-examination

A badly designed incrementality test produces numbers worse than last-click, because they look scientific. The discipline lives in the design, before the results.

1
Define the minimum detectable effect

Before the test you fix the MDE: the smallest lift that would justify a budget decision. From that and the baseline conversion rate you run the power analysis and size the sample.

2
Build the control group

Choose the design: geo holdout with matched pairs and synthetic control, or PSA/ghost ads for user-level causality. Use a pre-period equal to or longer than the test for a clean difference-in-differences read.

3
Isolate contamination

Police geographic spillover, audience overlap and leakage between groups. A contaminated control inflates or zeroes the lift and makes the test unusable, regardless of apparent significance.

4
Read the lift with its uncertainty

Report the effect with a confidence interval and statistical power, never a single number. Thresholds of 90–95% confidence and 80–90% power separate a real signal from seasonal noise.

Design and power analysis in days; typical test window 4–8 weeks plus an equivalent pre-period; read and calibration immediately after.

Comparison · Methods

Which method for which question

There is no single winner. Each method answers one question well and the others poorly: the choice depends on what you must decide and what granularity you have.

MethodAnswersLimit to know
Last-click / multi-touchWhich paths touched the conversionDoes not estimate the counterfactual, over-credits the lower funnel, depends on trackable identity
Geo holdout / lift testThe causal market-level effect of a channelNeeds enough markets, a pre-period and spillover control; answers slowly
PSA / ghost adsUser-level causality inside a walled gardenCostly (control spend) and poorly portable across platforms
Marketing mix modelingStrategic portfolio allocation, incl. offline and brandAggregate and slow; suffers multicollinearity without external calibration
Attention (post-viewability)Whether the ad was actually noticed, not just visibleMeasures exposure quality, not the sale: must be tied back to outcome
System · One surface

Triangulation on a single read

The operational output is not a method, it is a system where each component feeds the next and produces one defensible budget decision.

01

Experiment as causal anchor

Geo tests and holdouts provide ground truth: the one fixed point that does not depend on model assumptions or cookies.

02

MMM calibrated on experiments

Experiment results enter the marketing mix model as Bayesian priors: the model treats them as given, corrects premature saturation, and optimal allocation shifts accordingly.

03

Attention above viewability

The attention layer uses MRC viewability as a baseline and adds attentive seconds: it separates inventory that makes the message work from inventory that wastes it.

04

Attribution as local diagnostic

Attribution stays useful for the granular detail of creative and tactic, once anchored to causal lift instead of used as absolute truth.

A campaign worked when, turning it off, the business goes down. Everything else is bookkeeping.

Questions · Decisions

The questions an enterprise buyer asks

Can we keep using last-click?

Yes, as granular diagnostics, never as truth for deciding budget. Anchor it to incrementality experiments: without a counterfactual it systematically over-credits channels that intercept demand that already exists.

What does a credible incrementality test take?

Design matters more than duration. You need a minimum detectable effect and power analysis before you start, a clean control group, a pre-period at least as long as the test, and active contamination control. A typical window is 4–8 weeks plus pre-period.

Are attention and viewability the same thing?

No. MRC viewability certifies the opportunity to be seen (50% of pixels for 1 second for display, 2 seconds for video). Attention, per the 2025 IAB/MRC guidelines, starts from that baseline and measures whether the ad was actually noticed.

Cases

From problem to result — anonymised.

Retail · anonymised

The channel you pause while sales hold

Problem A multinational retailer kept scaling spend on brand search and prospecting display, trusting last-click, which reported them as the most efficient channels in the mix.

Method Geo holdout with matched-pair markets and a synthetic control, a pre-period as long as the test, active spillover control across regions, and the lift read with its confidence interval.

Result Part of the lower-funnel spend proved non-incremental — it intercepted demand that converted anyway — and the freed budget was shifted to channels with proven causal lift, with no loss of revenue.

CPG · anonymised

The MMM that couldn't see saturation

Problem A consumer-goods manufacturer ran a marketing mix model that suffered multicollinearity and kept recommending spend on an already saturated channel, while leadership could not defend the allocation to the board.

Method Causal lift tests on the key channels, then calibrating the model by treating the experiment results as Bayesian priors, so the MMM corrected premature saturation and recomputed optimal allocation.

Result Portfolio allocation shifted toward levers with real remaining marginal return, and the calibrated model became the instrument for defending budget with causal proof instead of charts that go up.

Financial services · anonymised

High viewability, low recall

Problem A financial-services brand bought premium inventory with very high certified viewability, yet brand tracking showed disappointing ad recall and no one could explain the gap.

Method Adding an attention layer above the MRC viewability baseline per the IAB/MRC guidelines, measuring attentive seconds and tying exposure back to the brand outcome instead of stopping at the opportunity to be seen.

Result Technically visible but barely noticed inventory was separated from genuinely attentive placements, and the plan was re-pointed toward positions that made the message work rather than just serving it.

Go deeper

Bring this to your stack.