Advanced Attribution Modelling · Multi-Touch, Incrementality & MMM

Last-click attribution assigns 100% of conversion credit to the final touchpoint before purchase. It is the default in most analytics platforms, most ad platform reports, and most marketing dashboards. It is also systematically wrong in predictable ways that distort marketing investment decisions.

The failure modes are well-documented: it over-credits bottom-funnel channels (paid brand search, retargeting, cashback affiliates) that intercept purchase intent created by upper-funnel channels; it under-credits awareness and consideration channels (display, social, content) that initiate the purchase journey; and it creates perverse incentives — marketers optimising for last-click ROAS will defund the channels that create the demand their bottom-funnel channels harvest.

The classic documented symptom: a brand cuts its display and social budget to improve blended ROAS, and for three weeks ROAS improves. Then paid search volume starts declining because there are fewer in-market prospects. By month two the full impact of reduced upper-funnel investment is visible — but attribution models gave no warning because they never credited upper-funnel channels with the conversions they were driving.

Rule-based models apply fixed credit distribution logic regardless of the actual causal relationship between touchpoints and conversion:

All rule-based models share a fundamental limitation: they distribute credit across observed touchpoints without knowing which touchpoints were causally necessary. A customer who would have converted regardless of seeing a display ad is still crediting the display channel in every model. This is why model comparison — switching from last-click to data-driven — does not solve attribution; it only changes how you distribute credit across the same flawed measurement universe.

Google's Data-Driven Attribution (DDA), available in GA4 and Google Ads, uses machine learning to assign conversion credit based on the actual contribution of each touchpoint across all observed conversion paths. Instead of applying fixed rules, DDA compares paths that converted with similar paths that did not convert to estimate how much each touchpoint changed the probability of conversion.

The technical mechanism is a variant of the Shapley value from cooperative game theory — a Nobel Prize-winning framework for fairly distributing credit among participants in a cooperative game. Each channel's Shapley value represents its marginal contribution to conversion probability across all possible orderings of the touchpoints in the path.

DDA's documented advantages over rule-based models: it adapts to actual customer journey patterns rather than applying fixed weights; it reflects real touchpoint interactions (some channels have synergistic effects; others have diminishing returns); and it updates as customer behaviour changes.

DDA's documented limitations: it is correlation-based, not causal — a touchpoint that consistently appears on paths that convert will receive high credit even if it is not causally necessary; it requires a minimum of ~800 conversions to train (Google's documented minimum); it operates only within the Google ecosystem and cannot account for touchpoints outside it; and it cannot resolve cross-device attribution gaps.

Incrementality testing directly measures the causal effect of advertising by randomly withholding it from a control group and measuring the difference in conversion rate versus an exposed treatment group. Unlike attribution models — which distribute credit across observed touchpoints — incrementality testing answers the causal question: would this conversion have happened without this channel?

The experimental design: define a target audience; randomly split it into treatment (receives advertising as normal) and holdout (advertising withheld); run for a statistically sufficient period; compare conversion rates. The incremental lift = (treatment conversion rate − holdout conversion rate) / holdout conversion rate.

Key design decisions that determine test validity:

Sample size and statistical power: The test needs sufficient conversions in both groups to detect the expected effect size at a meaningful confidence level. A 5% expected incremental lift requires a larger sample than a 30% lift. Calculate minimum detectable effect and required sample size before designing the test — not after.

Holdout isolation: The holdout group must genuinely receive no advertising from the tested channel. Leakage — where holdout users are inadvertently exposed through shared device IDs, cookies, or IP addresses — inflates the holdout's conversion rate and understates incremental lift.

Test duration: Tests must run for at least two full purchase cycle lengths. A category with a 30-day consideration cycle needs a test running 60+ days to capture the full incremental effect. Tests that are too short systematically understate lift for long-cycle categories.

User-level holdout testing — withholding ads from individual users — is difficult to execute cleanly because users cross devices, platforms track imperfectly, and platform APIs often do not support true exclusion at scale. Geo-based holdout experiments address this by using geographic regions as the experimental unit rather than individual users.

The design: identify matched pairs of geographic regions with similar baseline conversion rates and demographic profiles; run advertising normally in treatment regions; suppress advertising in holdout regions; compare conversion rates between matched pairs. Because all users in a holdout region receive no advertising (without any individual-level exclusion), there is no leakage problem.

Google's Causal Impact and Meta's GeoLift are documented open-source tools for geo experiment design and analysis. Both use synthetic control methods — constructing a statistical counterfactual from the holdout regions to estimate what the treatment regions' conversions would have been without advertising.

Geo experiments are the most rigorous incrementality testing available for most organisations because they avoid individual-level tracking challenges. Their limitation: they require meaningful geographic variation in spending and sufficient conversion volume at the regional level to produce statistically reliable results — not always feasible for small businesses or narrow geographic markets.

Marketing Mix Modelling (MMM) is a statistical technique that decomposes observed business outcomes (revenue, sales volume) into contributions from marketing channels and non-marketing factors (seasonality, price changes, macroeconomic conditions, distribution changes). Unlike attribution and incrementality testing — which operate at the user or group level — MMM operates at the aggregate time-series level, using regression to identify channel contribution patterns.

MMM's documented advantages: it can include offline channels (TV, radio, outdoor) alongside digital; it captures long-term effects and carryover (adstock) — the documented phenomenon where advertising effects persist beyond the immediate period of exposure; it does not require user-level tracking data, making it GDPR-compatible and cookieless by design; and it provides a unified cross-channel view that attribution models cannot produce.

MMM's documented limitations: it requires 2–3 years of historical data to produce reliable models (shorter histories produce unstable coefficients); it is a lagging indicator — model results are available weeks or months after the period being analysed; and collinearity between channels (e.g., search and social spending typically both increase together at peak periods) makes individual channel coefficients unreliable without experimental variation in spending levels.

The complementary relationship between MMM and incrementality testing: MMM provides the strategic view — how channels contribute to revenue at the aggregate level over time. Incrementality testing provides the tactical view — whether a specific channel or campaign is generating incremental conversions now. The two methods produce different answers because they measure different things. Best-in-class measurement organisations use both, triangulating toward a coherent picture of channel contribution. See the dedicated Marketing Mix Modelling guide for implementation detail.

Every major advertising platform reports its own attributed conversions using its own attribution model — and every platform's self-reported numbers systematically overstate its contribution. This is not fraud; it is a measurement reality that follows from the structure of multi-channel advertising.

The documented mechanism: Meta and Google both use view-through attribution (crediting conversions that happen within a defined window after seeing an ad, even without a click). A user who sees a Meta ad, then searches on Google, then converts, will be counted as a Meta conversion (view-through) and a Google conversion (last-click). The sum of all platform-reported conversions exceeds the actual number of conversions — often by 2–5× for mature multi-channel advertisers.

Quantifying the inflation: run a controlled incrementality test on one platform while holding all other channels constant. Compare the incremental lift measured in the test to the platform's self-reported attributed conversions. The ratio tells you the inflation factor — documented research from large multi-channel advertisers consistently shows platform self-reported conversions running 1.5–3× above measured incremental conversions.

The practical implication: never use platform-reported ROAS as the definitive measure of channel efficiency. Use it as a relative signal within a platform (campaign A vs campaign B) while using GA4 data-driven attribution and periodic incrementality tests for absolute channel evaluation.

A unified measurement framework layers complementary methodologies to triangulate a reliable picture of marketing contribution:

When the three layers produce inconsistent answers — which they will — the reconciliation process is itself informative. A channel with high GA4 attribution, low incrementality, and small MMM coefficient is harvesting organic conversions at above-market cost. A channel with low GA4 attribution, high incrementality, and large MMM coefficient is generating demand that other channels are capturing credit for. Both situations require budget reallocation.

Attribution reform is as much an organisational challenge as a technical one. Channel managers whose bonuses depend on last-click ROAS have strong incentives to resist measurement changes that would reduce their reported numbers — even if those changes produce a more accurate picture of performance.

Common organisational resistance patterns: channel teams arguing against holdout testing ("we can't afford to withhold ads from any users"); media agencies producing attribution model comparisons that always show their channels performing well under their preferred model; and finance teams accepting platform-reported ROAS without question because it is the easiest number to get.

The governance change that most improves attribution practice: separating the teams responsible for running channels from the teams responsible for measuring them. When the team that runs Google Ads is also responsible for reporting Google Ads ROAS, the measurement will be unconsciously biased. When measurement is owned by a central analytics function, incentive-driven distortion is reduced.

Further Reading

Go deeper with these reference guides from the Digital Codex library.

Sources & References