What an attribution model decides
Every conversion has a history: the ads clicked, the searches made, the emails opened and the pages visited before it. AttributionAssigning conversion credit to touchpoints. is the act of assigning the value of that conversion back to the history, and a model is the rule that does the assigning. The rule decides three things: which touchpoints are eligible for credit at all, how the credit is split among them, and how far back in time the history is allowed to run, the lookback window.
As a framework, attribution is a lens rather than a measurement. Change the model and the same conversions produce a different ranking of channels, which means the model is a choice about what the organisation wants to see. Last click shows what closed; first click shows what introduced; linear shows participation; time decay shows recency; position-based shows the bookends; data-driven shows a statistical estimate of contribution within the tracked paths. The framework makes the choice explicit, which is its value, and it hides everything that was not tracked, which is its main danger. The attribution hub covers the measurement discipline as a whole; this page is about using a model as a thinking tool. How a specific tool applies these rules is a separate matter, and the attribution settings guide for Google Analytics covers the reporting model, the lookback windows and the channel groups that tool documents.
The models
The models fall into two families. Rule-based models apply the same split to every path regardless of what the data says; they are simple, stable and arbitrary. Data-driven models use the conversion paths themselves to estimate how much each touchpoint changed the probability of converting; they are more defensible and less transparent. The table shows what each makes visible and what each hides.
| Model | How credit is assigned | What it makes visible | What it hides |
|---|---|---|---|
| Last click | All credit to the final tracked touchpoint | What closes; good for capture channels | Everything that created the demand |
| First click | All credit to the first tracked touchpoint | What introduces; good for discovery channels | Everything that nurtured and closed |
| Linear | Equal credit to every touchpoint | Participation across the path | Which touches actually mattered |
| Time decay | More credit the closer a touchpoint is to conversion | Recency; favours channels near the end | Early influence in long cycles |
| Position-based | Most credit split between first and last, the rest shared across the middle | Introduction and close together | The middle of the path, by design |
| Data-driven | Credit estimated from comparing converting and non-converting paths | Statistical contribution within tracked data | Its own method; a tool's model is only partly documented, and small datasets fall back to simpler rules |
The side-by-side detail, including the arithmetic of each rule and where each one misleads, is in the attribution models compared guide. The framework point is simpler: every model is a statement about which part of the path the organisation has decided to reward. Google Analytics offers only some of these in its reporting settings; the settings guide linked above covers which, and the help centre cited below is the documented reference for what each does.
Choosing a model for a decision
The model follows the decision, not the other way round. A budget decision about capture channels is served by last click, because those channels are judged on closing. A decision about whether a content programme or a video campaign is earning its place needs a model that credits early touches, or the programme will be defunded for doing exactly its job. A reporting standard that the whole company uses should be the model with the fewest obvious biases for the mix in question, which is often data-driven where the volume supports it and position-based where it does not, with the choice written down and dated.
Three questions settle most cases. What is being decided, and by whom? Which touchpoints does that decision need to see fairly? Is there enough conversion volume for a data-driven estimate to be stable, or will the tool quietly fall back to a rule? The choosing an attribution approach guide turns those questions into a decision framework that also brings in incrementality and mix modelling, and the measurement plan framework is where the chosen model is recorded next to the CVRShare of visits or clicks that convert. and other metrics it will be applied to. Changing the model without re-stating the targets built on the old one is the most common way an attribution choice turns into an argument.
Worked path
The scenario below runs one tracked path through each model to show how the ranking of channels changes with the lens. The figures are invented.
One purchase, six models
A customer's tracked path to a £120 order over 14 days: an Instagram ad click, then an organic search visit three days later, then a click on a promotional email on day 10, then a direct visit that converts on day 14. Under last click the direct visit receives the full £120, or under a last non-direct rule the email does. First click gives Instagram £120. Linear gives each of the four touchpoints £30. Time decay with a short half-life might give direct £55, email £35, search £20 and Instagram £10. Position-based gives Instagram £48 and direct £48, with search and email sharing the remaining £24. A data-driven model, trained on 9,000 paths from the same site, estimates that paths including the email convert at 2.1 times the rate of otherwise similar paths without it, and assigns email £60, Instagram £30, search £18 and direct £12. Six models, six rankings of the same four channels. The team's decision was about whether to keep the email programme; data-driven and time decay both said yes, last click said no, and the choice of model decided the budget.
Limits
Attribution models share one limit that dwarfs the differences between them: they divide credit among tracked touchpoints only. An ad seen but not clicked, a podcast mention, a recommendation from a friend, a visit in a browser where consent was refused, a purchase in a store: none appears in the path, so none receives credit, and the tracked channels absorb value they did not create. Cross-device journeysA journey in which the ad interaction and the conversion happen on different devices, so that only a signed-in identity or modelling can connect them. fragment into separate paths unless the person is logged in. Lookback windows cut long cycles short. Walled-garden platforms report their own attribution on their own terms, and the numbers rarely reconcile with the analytics tool's.
Data-driven models add limits of their own. The method is proprietary and only partly documented, so the organisation is trusting a result it cannot inspect. They need volume; below the documented threshold the tool falls back to rules, and not always prominently. And they measure contribution within the tracked paths, which is not the same as the incremental sales the channel caused; a channel that only reaches people who were going to buy anyway can look excellent in a data-driven model.
Pairing with incrementality and mix modelling
Because attribution sees only tracked paths, two other methods are usually paired with it. IncrementalityConversions caused by marketing. testing withholds a channel from a comparable group and measures the difference in outcomes, which answers the causal question attribution cannot: did this channel produce sales that would not otherwise have happened? Marketing mix modelling, MMMStatistical model of channel impact., uses statistical models of aggregate spend and sales over time, including untracked channels and external factors, to estimate each channel's contribution without any user-level tracking.
The three fit together by timescale and granularity. Attribution is daily and granular and biased towards tracked, late-path channels; it is the operating report. Incrementality tests are occasional and precise; they calibrate attribution for the channels that matter most. Mix modelling is quarterly and coarse; it sets the frame within which the other two are read. A team that uses attribution alone optimises towards whatever its model rewards; one that adds the other two can tell the difference between a channel that closes and a channel that causes.
Common mistakes
- Asking which model is right. None is; ask which decision the model is for.
- Switching models without re-basing targets. The targets were set under the old lens; the new one changes every channel's apparent performance.
- Reading data-driven as causal. It estimates contribution within tracked paths, not incremental sales.
- Forgetting the untracked. Offline, view-through and consent-refused journeys never enter the model.
- Trusting platform-reported attribution over a single standard. Each platform credits itself; one model in one tool should be the reference.
- Reporting at the lookback window the tool defaulted to. The window is a decision, and it should match the sales cycle.
