Attribution & Measurement

Performance marketing is held to a number — supplied by the platforms selling you the media. So the discipline lives or dies on measurement, and no single method tells the truth: attribution is fast but biased, experiments are slow but causal, MMM is strategic but top-down. This module is the triangle — how to steer with attribution, prove with experiments, and plan with MMM, and which one wins when they disagree.

Why measurement is the whole game

Measurement is the whole game in performance marketing, because the discipline is held to a number — and the number is supplied by the platforms selling you the media. No single method tells the truth: attribution is fast but biased, experiments are slow but causal, and MMM is strategic but top-down. You triangulate the three; you never trust one corner alone. Get this wrong and every other decision is optimized toward a fiction.

Every module so far has pointed here. We said attribution flatters (module 1), that the machine optimizes toward whatever number you feed it (module 3), that a creative “winner” must be re-proven (module 5). All of those warnings are really one warning: the measurement is the foundation, and most of the industry is building on a cracked one. This module is how to pour a foundation you can actually stand on.

The central image is a triangle, not a single dashboard. Attribution, incrementality experiments, and marketing-mix modeling each answer a different question at a different speed, and the skill is knowing which corner to ask. Trusting any one of them alone is the mistake that cost Adidas four years and Uber a hundred million dollars — both stories from module 1 were, at root, measurement failures.

What attribution is — and is not

Attribution assigns credit for conversions to the touchpoints a user was observed to encounter. It is correlational, not causal: it can tell you which ads a converter saw, never whether those ads caused the conversion. That single distinction — observed correlation versus proven cause — is the most important idea in marketing measurement, and the reason attribution alone can never set a budget.

Start by being precise about what attribution actually does, because the word gets used as if it meant “truth” and it does not. Attribution observes the touchpoints a person encountered on the way to converting and divides credit among them by some rule. Notice what is missing from that sentence: any information about what the person would have done without the ads. That counterfactual — the road not taken — is the only thing that proves causation, and attribution is structurally blind to it.

The attribution models and their biases

Attribution models differ only in how they divide credit among observed touchpoints — and each rule bakes in a bias. Last-click over-credits the bottom of the funnel, first-click the top, linear and time-decay spread credit by a rule you invented, and data-driven assigns algorithmic credit from observed paths. Data-driven is the least-bad for steering, but all of them are correlational and none can set a budget.

There is no “correct” attribution model, only models that are wrong in different, predictable directions. Knowing the direction of each bias is what lets you use a model without being fooled by it. Tap through the four you will actually meet and note the move each one demands.

Attribution as compass, not courtroom

Use attribution as a compass, not a courtroom: it is the right signal for steering daily optimization — pausing losers, nudging bids — and the wrong instrument for the verdict of how much to spend on each channel. The escalation ladder is simple: attribution to steer, experiments to prove, MMM to plan. Match the question to the corner of the triangle.

The practical art of measurement is matching the question to the method, because each method is excellent at one job and dangerous at the others. Attribution is fast and granular — perfect for the thousand small daily decisions — and biased and correlational, which disqualifies it from the few big ones. The ladder below routes each question to the tool that can actually answer it.

The measurement floor: server-side and one primary action

The measurement floor is the infrastructure you install before spending a dollar: server-side conversion tracking so signal survives ATT and ad-blockers, exactly one primary conversion action per goal, disciplined UTMs, and an MER baseline from historicals. Skip the floor and you spend a quarter optimizing toward a number that was wrong from day one.

Before any of the clever measurement methods matter, you need a measurement floor — the unglamorous plumbing that determines whether the data flowing into all three corners of the triangle is even trustworthy. Post-ATT, the most consequential piece is server-side conversion tracking, because client-side pixels now miss a large and growing share of conversions to tracking prevention and ad-blockers.

Claim: Server-side conversion tracking recovers conversion signal that client-side pixels lose to ATT, ad-blockers, and browser tracking prevention — a gap that can reach 30%+ of conversions on affected traffic. Source: RGM analysis from client measurement audits. Context: Without server-side signal, the data feeding your attribution, experiments, AND bidding is missing a non-random chunk — biasing every downstream decision.

The reconciliation: claimed vs true

Reconciliation is the act of correcting a platform-claimed number into a believable one: true incremental ROAS equals the claimed ROAS multiplied by measured incrementality, judged against your breakeven (1 ÷ margin). This is the same flattery correction as module 1, but here producing the incrementality number is measurement’s explicit job — via experiments, not assumptions.

Module 1 introduced the flattery correction as a way to read a dashboard skeptically. Here it becomes an operating procedure with a number you actually measure rather than estimate. Drag the sliders with your real figures: claimed return, the incrementality a holdout gave you, your margin. Watch a celebrated 5× campaign cross below breakeven the moment the measured truth enters.

Incrementality: the only proof of cause

Incrementality is the only method that proves causation, because it manufactures the counterfactual attribution lacks: withhold ads from a matched control group, run them in a test group, and the difference in outcomes is the causal effect. Geo holdouts, conversion-lift studies, and ghost/PSA tests are the main designs. The delta between test and control — not any attributed number — is what sets a channel’s real budget.

If attribution is the compass, incrementality is the act of actually walking the territory to see where the compass lied. The mechanism is simple and almost impossible to argue with: change the advertising in some markets and not others, keep everything else matched, and the revenue gap can only be explained by the advertising. It is slow and it costs some volume, which is exactly why almost nobody runs it voluntarily — and exactly why the teams that do find the phantom ROAS everyone else is funding.

MMM: planning the year, and its open-source reset

Marketing mix modeling plans the year from the top down: it regresses sales against all marketing spend plus external factors to estimate each channel’s contribution and saturation, using no individual tracking at all — which is why it survives the cookieless era. Once a $50K-500K consulting engagement, it is now free and open-source (Meta’s Robyn, Google’s Meridian). It must be calibrated with experiments or it is only confident correlation.

The third corner of the triangle is the oldest and, for planning, the most powerful. MMM never looked at individuals, so privacy changes did not touch it — and as user-level attribution decayed, a 1960s technique came roaring back, now democratized by free Bayesian engines. The catch is the part the tutorials skip: a model fed only observational data produces precise-looking numbers that are pure correlation. Calibration with experiments is what turns it from a persuasive chart into a planning instrument.

Advanced playbook

Advanced measurement is triangulation as a standing operating system: attribution steering daily, a quarterly incrementality calendar proving causation channel by channel, and a calibrated MMM planning the annual mix — reconciled on a fixed cadence, with the experiment winning every disagreement. The outputs that reach finance are MER and incrementality-corrected iROAS, never platform ROAS alone.

The senior move is to stop treating measurement as a reporting function and start running it as a system with a cadence. Each corner of the triangle has a job and a schedule; the discipline is reconciling them on purpose and letting the most rigorous method — the experiment — settle disputes, then re-anchoring the others to it. The build below is how that system goes up.

Common mistakes

The classic measurement mistakes share one root: trusting one corner of the triangle as if it were the whole truth. Budgeting off attribution, never running a holdout, shipping an uncalibrated MMM, optimizing a metric until it stops measuring the business, and deferring the measurement floor are the recurring five.

• Setting budgets off platform attribution. It is correlational and flatters bottom-funnel channels; budget belongs on incrementality, not on credit the platform assigned itself.
• Never running an incrementality test. Without a holdout you are guessing at causation forever — and over-funding the channels attribution loves (ask Uber).
• Shipping an uncalibrated MMM. Free engines made MMM easy; unanchored, it launders correlation into false confidence. Calibrate with experiments.
• Optimizing a metric into meaninglessness. Goodhart’s Law: push a proxy hard enough and the machine games it. Optimize the business outcome, validate the proxy.
• Deferring the measurement floor. “Sort out tracking later” means a quarter of decisions built on missing, biased signal. Floor first, spend second.
• Reporting platform ROAS to finance. It is the gameable numerator; MER and incrementality-corrected iROAS are the numbers a CFO can trust.

Quick answers

What is the difference between attribution and incrementality?

Attribution assigns credit for a conversion to touchpoints the user was observed to interact with — it is fast but correlational, and it cannot see what would have happened without the ad. Incrementality measures causation: by withholding ads from a matched control group (a holdout) and comparing outcomes, it reveals how many conversions the advertising actually caused. Attribution steers daily decisions; incrementality decides budgets.

Why is multi-touch attribution declining?

Multi-touch attribution depended on tracking individual users across sites and apps, which privacy changes broke: Apple’s App Tracking Transparency cut iOS opt-in to roughly 14%, third-party cookies are disappearing, and platforms have removed view-through windows. With the user-level data gone, MTA can no longer see most of the journey, so the industry is shifting to aggregate methods (MMM) and causal methods (experiments) that never needed cookies.

What is marketing mix modeling (MMM)?

MMM is a top-down statistical method that regresses sales against marketing spend across all channels plus external factors (seasonality, price, promotions) to estimate each channel’s contribution and its saturation point — with no individual user tracking at all. Once a half-million-dollar consulting engagement, it is now available as free open-source Bayesian engines: Meta’s Robyn (2020) and Google’s Meridian (2024). It plans the annual budget; it should be calibrated with experiments.

Which attribution model should I use?

For daily steering, data-driven attribution is the least-bad model — but no attribution model should set your budget, because all of them are correlational and inside the walled garden. Use attribution as a compass for in-platform optimization, then validate the channels it flatters (especially brand search and retargeting) with incrementality experiments before allocating real budget to them.

How do I measure marketing in a cookieless world?

Triangulate three methods, each answering a different question: data-driven attribution to steer day to day (fast, biased), incrementality experiments such as geo holdouts to prove what each channel actually causes (slow, unbiased), and marketing mix modeling to plan the annual budget and find saturation (top-down, no tracking). Where they disagree, the experiment wins and the others are re-anchored to it.

What is a geo holdout test?

A geo holdout is an incrementality experiment where you turn a channel off (or change its spend) in a set of matched geographic markets while keeping it running in comparable control markets, then compare the revenue difference. Because the only thing that changed between the test and control geos is the advertising, the gap is a causal estimate of that channel’s true incremental return — the counterfactual that attribution can never provide.

Operating checklist — score yourself

Use this as the operating standard for marketing measurement in the cookieless era. None of it is a single dashboard — it is the discipline of triangulating three imperfect methods so the number you bet a budget on is one you actually proved.