Attribution Model

Origin & evolution

For most of the web's history the question answered itself by default. Analytics tools recorded the last interaction before a conversion and gave it the credit, because that was the click the system could see at the moment of purchase. The model was rarely chosen; it was inherited from how the data happened to be collected.

The turn came when Google Analytics introduced Multi-Channel Funnels on 24 August 2011, exposing the full sequence of channels a customer touched before converting and offering a menu of models over that sequence: first-touch, last-touch, linear (split the credit evenly), time-decay (weight recent touches more), and position-based (load the first and last). Suddenly the assumption was visible and adjustable, and the industry discovered it had been crediting the wrong things for years. Data-driven attribution arrived next, using machine learning🧠LearningValidationAn insight gained from an experimentView reference → over thousands of journeys to allocate credit by observed pattern instead of a fixed rule.

Then the discipline ran into a wall it has not climbed. Every multi-touch model, the algorithmic ones included, describes correlation. It watches the touchpoints a converting customer happened to pass through and assigns credit by some rule, but it never runs the counterfactual: would this customer have converted anyway, without that touch? Practitioners now state this plainly, and signal loss from privacy changes and cookie deprecation made cross-device journey stitching shakier still. The field's answer has been to move toward methods that test cause directly: incrementality experiments🧫ExperimentValidationA test designed to validate a hypothesisView reference → that hold out a control group, and marketing mix modelling (MMM), the regression approach that estimates each channel's contribution to sales at the aggregate without needing to track individuals.

How it works in practice

A subscription🔂SubscriptionSales & RevenueA recurring subscriptionView reference → app spends 300,000 dollars a month across paid search, paid social, and a podcast sponsorship. Under last-touch, paid search takes 70 per cent of the credit, because people Google the brand name right before signing up. The growth lead nearly cuts the podcast, which last-touch shows driving almost nothing.

Before doing so the team runs a geo holdout: they pause the podcast in three matched regions for six weeks and watch signups against untouched control regions. Signups in the paused regions fall 12 per cent, far more than any attribution model credited the channel. Branded search was harvesting demand the podcast created; last-touch saw the harvest and missed the planting. The team keeps the podcast and reweights the attribution model to inform daily bidding while treating the holdout result, not the model, as the truth about incremental impact.

Attribution model vs. its neighbours

• Funnel⬇️FunnelGrowthA conversion funnel tracking user progressionView reference → describes the stages a customer moves through on the way to converting. An attribution model assigns credit across the touchpoints inside that journey. The funnel is the map; attribution is the accounting laid over it.
• Acquisition channel🎣Acquisition ChannelGrowthA channel through which users are acquiredView reference → is one of the things an attribution model distributes credit among. A channel is a place a touch happens; the model decides how much that touch was worth, and the same channel looks heroic or worthless depending on which model you choose.
• Metric🔢MetricStrategyA unified metric that measures progress, health, or behaviour across the productView reference → is a measured quantity, conversions or revenue, that exists whether or not anyone attributes it. An attribution model is the allocation logic that splits a metric across causes, and it can be wrong while every underlying metric stays correct.

In the graph

In the Unified Product Graph, an attribution model sits in the growth region, joined to a product by product_attributed_via_attribution_modelProductattributed viaAttribution Modelhierarchy. Recording the model as an explicit entity rather than an invisible default makes the assumption inspectable, which is the whole point. A product attributed solely via a single last-touch model, with no connected incrementality experiment or mix model anywhere in the graph, is a product whose growth decisions⚖️DecisionStrategyA recorded decision with context, rationale, and consequencesView reference → rest on correlation it has never tested, and the structure makes that exposure visible instead of buried in a dashboard🎛️DashboardData & AnalyticsAn analytics dashboardView reference →'s default setting.