Credit is not causation, and most marketing reporting measures credit.

The two questions, stated precisely

Attribution asks: given that a conversion happened, which touchpoints should receive credit and in what proportion? Incrementality asks: how many of these conversions would not have happened without the spend? The first is a bookkeeping convention applied to observed paths. The second is a counterfactual, and counterfactuals are not observable — they have to be estimated from a group that did not receive the treatment.

This is why attribution always returns a number and incrementality sometimes cannot. Attribution takes conversions that occurred and distributes them according to a rule; there is no version of that computation that fails. Incrementality requires withholding spend from a comparable group and comparing outcomes, and if the design is not feasible, the honest output is that the answer is unknown.

The practical consequence is that attributed return and incremental return can differ by a large factor in either direction, and routinely do. A channel that mostly reaches people who were already going to convert will show strong attributed performance and weak incrementality. A channel that generates demand which later converts through a different path will show the opposite.

Why last-click persists despite being wrong

Last-click gives all credit to the final touchpoint before conversion. As a causal model it is indefensible — it assigns everything to whatever the buyer touched last, which is frequently branded search, a channel largely harvesting demand created elsewhere. As an operational convention it survives because it is deterministic, cheap, universally understood and never disputed between two people looking at the same data.

That last property is worth taking seriously instead of dismissing. A measurement system that two teams can agree on has real organisational value even when it is biased, provided everyone understands what it is and is not saying. The damage comes from treating it as causal — cutting an upper-funnel channel because it shows little last-click return is exactly the mistake the model's structure guarantees.

Multi-touch attribution distributes credit across the path instead, either by rule (linear, time-decay, position-based) or by a fitted model. Rule-based versions replace one arbitrary convention with a slightly less arbitrary one; they are not more causal, just less concentrated. Data-driven versions fit weights from observed paths, which is a genuine improvement in describing correlation and still not an estimate of causation, because the buyers who saw an ad differ systematically from those who did not.

The observability problems attribution cannot model around

Every attribution model assumes it can see the path. Increasingly it cannot. Cross-device journeys break when someone researches on a phone and buys on a laptop. Walled gardens report conversions they attribute to themselves without exposing user-level paths for independent reconciliation. Offline conversions, phone orders and retail purchases leave no digital trail at all. View-through effects are invisible by construction.

These are not tuning problems. A model cannot allocate credit across touchpoints it never observed, so the touchpoints it did observe absorb the credit — which systematically over-credits the measurable channels and under-credits everything else. The bias has a direction, and the direction favours whatever is easiest to track.

Platform self-reporting compounds it. When each platform reports conversions using its own attribution window and its own view-through rules, summing them across platforms reliably exceeds total conversions, because the same conversion is claimed more than once. That arithmetic alone should prevent platform-reported return being used for cross-channel budget decisions, and it frequently does not.

Identity resolution helps at the margin and does not solve it. A resolved customer identity joins more of the path than a cookie does, particularly for logged-in and offline activity, and it still cannot observe the impression somebody saw and did not click.

• Never sum platform-reported conversions across platforms; the same conversion is counted more than once
• State the attribution window explicitly — changing it changes every channel's apparent return
• Treat untracked channels as unmeasured rather than as underperforming
• Expect the bias to favour whatever is easiest to track, not whatever works

How incrementality is actually established

Incrementality requires a comparison group that did not receive the spend and is otherwise similar. The cleanest form is a randomised holdout: a share of the audience is excluded from targeting, and the outcome difference between groups estimates the causal effect. Where individual-level holdouts are impossible, geography is the usual substitute.

Geo experiments randomise regions rather than people. Google's published methodology describes the shape directly: non-overlapping geographic regions are randomly assigned to control or treatment, each region realises its assigned condition through geo-targeted advertising, and the difference in outcomes estimates incremental effect. The authors' stated advantages are that the design is conceptually simple and the results are straightforward to interpret.

The reason geo works where user-level holdouts do not is that it sidesteps identity entirely. It does not need to know who saw what; it needs regions to be comparable and the treatment to be containable within them. That also gives it a second advantage: it captures offline conversions, because outcomes are measured as total regional sales rather than as tracked events.

The cost is real and should be stated before anyone commits. An experiment needs enough regions to randomise meaningfully, a baseline period, a test period long enough to cover the purchase cycle, and the deliberate acceptance of forgone revenue in the control group. It also produces one estimate for one channel at one spend level at one point in time — not a reusable measurement system.

Selection bias, and why the model cannot fix it

Advertising is targeted, and targeting is the whole problem. Platforms deliberately show ads to people most likely to convert, which means the exposed group differs from the unexposed group in precisely the way that predicts the outcome. Any comparison of exposed against unexposed therefore measures targeting quality plus advertising effect, with no way to separate them observationally.

Retargeting is the clearest illustration. It shows ads to people who already visited the site and often already added to basket — a group with a high baseline conversion rate before any ad is served. Attributed return looks outstanding. Incremental return is frequently a fraction of it, because a large share of those conversions were going to happen regardless.

This is also why statistical controls do not rescue observational comparison. Adjusting for observed covariates handles the differences you measured; targeting algorithms optimise against signals you did not and often cannot observe. Randomisation is not a nicety here — it is the only mechanism that balances the unobserved differences too.

Where marketing mix modelling fits

MMM regresses aggregate outcomes against aggregate spend and external factors over time, needing no user-level tracking at all. That makes it robust to the identity and privacy constraints that break attribution, and useful for broad budget allocation across channels including offline ones. It is a third instrument, not a replacement for either of the other two.

Its limitations follow from being observational and aggregate. It infers effects from historical variation, so a channel whose spend has barely changed cannot be estimated well; it is sensitive to specification choices around adstock, saturation and controls; and it produces channel-level guidance on a monthly or quarterly rhythm instead of the campaign-level decisions attribution informs.

The arrangement that works in practice uses all three for what each is good at: attribution for operational, in-flight decisions; experiments for periodic ground truth on the channels where the stakes justify the cost; and MMM for strategic allocation, calibrated against the experiment results. The calibration is the part that makes it a system instead of three opinions — an MMM validated against a holdout is considerably more trustworthy than one fitted in isolation.

Which instrument answers which question

Most measurement arguments dissolve once the question is stated precisely, because attribution and incrementality are rarely both appropriate for the same decision. The table below is organised by the question rather than by the method, which is the order the decision actually arrives in.

The pattern across the rows is consistent: attribution is adequate when the decision is operational, reversible and within a channel; an experiment is required when the decision is about whether a channel or budget level is worth anything at all.

Running an experiment that produces a usable answer

Most marketing experiments fail for design reasons rather than analytical ones: the test is too short to cover the purchase cycle, the effect is too small to detect at the available scale, or the control group was contaminated by spillover. Deciding in advance what effect size would change a decision, and whether the design can detect it, prevents most of that.

Work backwards from the decision. If a 5% lift would not change the budget, do not design a test that can only detect 5%. If the purchase cycle is six weeks, a two-week test measures a fraction of the effect and will understate it. If the channel is a small share of total spend, the effect may be undetectable against normal variation regardless of how long the test runs — and knowing that beforehand saves the cost of running it.

Contamination is the geo-specific risk. Media that does not respect regional boundaries, customers who shop across regions, and national promotions running concurrently all leak treatment into control and bias the estimate toward zero. Check for concurrent activity before committing, because a national promotion during the test window can invalidate the whole thing.

Finally, an experiment estimates the effect of one channel at one spend level in one period. It does not generalise to a different budget, a different season or a different creative strategy. Treat results as having a shelf life and re-run periodically on the channels where the stakes justify it.

• Define the effect size that would change a decision, before designing the test
• Run for at least one full purchase cycle, not one reporting period
• Check for concurrent national activity that would contaminate the comparison
• Record the spend level and period — the result does not transfer to others

The honest limits of both

Attribution is a convention that cannot establish causation no matter how sophisticated the model. Incrementality is causal but expensive, slow, narrow in scope and perishable. Neither gives a marketing team a continuously accurate, channel-level view of what its spend is doing, and any vendor claiming otherwise is describing attribution with better presentation.

What a mature setup looks like is therefore unglamorous: attribution used operationally with its bias understood and stated, a small number of experiments run each year on the channels where the stakes are highest, and those results used to calibrate the attributed numbers everyone reads day to day. The calibration factor is the useful artefact — knowing that a channel's attributed return is roughly twice its incremental return is enough to make better decisions with the reporting you already have.

Two things this does not cover. Long-term brand effects fall largely outside both instruments, because the effect horizon exceeds any practical test window and MMM's estimates of it are highly specification-dependent. And none of it addresses whether the conversion being optimised is the right outcome — a channel that drives volumes of low-value, high-churn customers can look strong on every measure here and lose money, which is a customer value question instead of a measurement one.

Questions and trade-offs

Technical references

• Vaver & Koehler (Google): Measuring Ad Effectiveness Using Geo Experiments ↗
• Jin et al. (Google): Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects ↗
• Google Ads Help: about attribution models ↗