Quick definition: An attribution model is the documented rule or algorithm that distributes credit for a conversion across recorded marketing or product touchpoints.
What is an attribution model?
An attribution model turns a sequence of recorded interactions into a credit allocation. If a person sees a display ad, clicks a search result, opens an email, and purchases, the model specifies whether one interaction gets all the credit or whether several share it. The same events can produce very different channel reports under different models, even though the observed journey has not changed.
Models are primarily descriptive tools. They provide a repeatable answer to “Where did our reporting system assign this conversion?” They do not automatically answer “What caused this conversion?” A more complicated model, including one trained on historical data, can still learn patterns of exposure and selection rather than causal effects. Model output should be presented as model-based credit, not as proof of incremental business value.
A useful model is one that matches a defined decision, has stable inputs, is understandable to the people using it, and has known limitations. It is not necessarily the model that assigns the largest amount of revenue to a favored channel.
Definitions and boundaries
A model is the credit-allocation rule. An attribution window is the time interval after an eligible touchpoint in which a conversion may receive credit. An identity rule determines when events are considered to belong to the same person, account, device, or household. A conversion definition specifies the outcome and its value.
All four elements are part of the measurement system. Calling a report “last click” without stating its window, identity resolution, and conversion logic leaves essential ambiguity. For example, a 30-day click window can credit an early search interaction that a 7-day window would exclude; merging a logged-in device with an anonymous browser can change both the path and the conversion count.
| Model | Credit rule | Useful descriptive question |
|---|---|---|
| First-click | All credit to first eligible touchpoint | Which recorded source initiated paths? |
| Last-click | All credit to final eligible touchpoint | Which source was last observed before conversion? |
| Linear | Equal shares across eligible touchpoints | Which sources appeared in recorded paths? |
| Time-decay | More weight to recent touchpoints | Which late-path interactions were recorded? |
| Position-based | Weighted shares for first, last, and middle | How does a chosen journey narrative distribute credit? |
Data-driven approaches estimate weights from historical paths. Their assumptions, training population, available features, and handling of unobserved exposures should be reviewed. They can summarize complex patterns, but they do not eliminate the need for causal experimentation.
Concrete scenario: choosing a model for a subscription funnel
A software company sells an annual plan. A prospect first arrives via a partner link, later clicks a paid-search ad, attends a webinar after an email invitation, and converts 18 days after the partner visit. The growth team wants a monthly channel report and must select a model before the quarter begins.
With a 30-day lookback, first-click gives the partner full credit. Last-click gives the webinar email full credit if it was the final eligible interaction. Linear spreads the credit across the partner, paid search, and email. A time-decay model favors the later interactions. None is the uniquely true answer; each encodes a different reporting convention.
The team chooses a simple, stable model for monthly operational reporting and shows a second view of assisted paths for investigation. It does not use either report alone to decide the paid-search budget. Instead, it reserves a randomized geographic or audience holdout to compare total qualified subscriptions with and without the paid-search campaign. That test estimates whether the campaign generated additional subscriptions, including those whose final recorded touchpoint was the email.
A model selection framework
- Write the decision first. Clarify whether the report supports channel operations, journey research, finance reconciliation, or a causal investment decision.
- Specify the measurement contract. Define conversions, eligible touchpoints, identity logic, time zone, revenue treatment, deduplication, and data-refresh policy.
- Choose windows from the buying cycle. Consider the typical delay from interaction to conversion and any credible response evidence. Avoid inherited defaults that do not fit the product.
- Use the simplest adequate rule. A rule that stakeholders can reproduce and explain is often better for routine reporting than an opaque model with uncertain inputs.
- Perform sensitivity analysis. Compare reasonable windows and models to identify decisions that reverse solely because of a convention.
- Set governance. Version the model, record owners and change dates, communicate breaks in trend, and do not revise the rule retroactively to favor a channel.
- Validate causal claims separately. Use experiments or other credible designs when a decision depends on incremental impact.
For a conversion value V and eligible touchpoints 1 through n, any model assigns weights wi where Σwi = 1. Credit for touchpoint i is V × wi. In a linear model, each weight is 1/n. This is arithmetic allocation, not a causal equation: the formula has no term for what would have happened had a touchpoint been absent.
Windows, identity, and data quality
Window selection has material effects. A long window may associate early awareness activity with later purchases that have many intervening influences. A short window may omit plausible delayed responses. View-through windows need special caution because an impression does not show attention or intent, and impressions can be numerous. Use different windows for clicks and views only when the definitions are explicit and consistently applied.
Identity resolution is another boundary, not a hidden implementation detail. Browser restrictions, consent choices, shared devices, account logins, cookie deletion, and cross-device behavior all change what a system can connect. A model describes the observed, linkable path—not the entire path a person experienced. Track match rate, duplicate rates, delayed conversion imports, and the fraction of conversions with no attributed touchpoint.
Make measurement quality visible beside credited revenue. A sudden rise in direct conversions may reflect a tag change rather than customer behavior. A channel’s decline may be caused by a shortened window, loss of a device identifier, or a vendor import failure. Before interpreting a trend, check the event pipeline and definitions.
Attribution models and experimentation
Experiments answer a different question from attribution models. In a randomized holdout, eligible units are assigned to receive or not receive an intervention. The difference in a pre-specified outcome estimates incremental effect under the design assumptions. An attribution model can be applied consistently to both groups for diagnostic reporting, but it must not replace the outcome comparison.
Keep the model, windows, identity logic, and conversion definition the same for treatment and control. Changing a window after looking at results, or allowing treatment to alter which events qualify, can bias the comparison. Report total causal outcomes and relevant guardrails before reporting attributed breakdowns. For a refresher on experimental controls, see how to write an A/B test hypothesis and the glossary page on A/B testing.
Attribution can still help explain an experiment. It may reveal that a treatment increased search follow-up or shifted the distribution of last touches. Treat such findings as mechanism hypotheses unless the design and analysis were built to estimate them.
Pitfalls and limitations
- Picking the model after results are known. This turns a reporting convention into an optimization target and undermines comparability.
- Comparing tools without matching definitions. Different vendors may use different windows, identities, and deduplication rules.
- Interpreting algorithmic weights causally. Historical correlations can reflect targeting, prior intent, or missing touchpoints.
- Forgetting data gaps. Consent, offline sales, cross-device loss, and blocked tracking affect who is represented.
- Using one report for every decision. Operational allocation, finance reconciliation, and causal investment evaluation require different evidence.
FAQ
What is the best attribution model?
There is no universal best model. Select a documented rule that fits the descriptive question and pair it with causal evidence for high-stakes spending decisions.
Should we use first-click or last-click?
Use either only for the question it answers: first-click describes recorded entry points, while last-click describes final recorded interactions. Neither proves causal credit.
How should we choose an attribution window?
Base it on the purchase cycle and plausible response timing, disclose it, and test whether conclusions change under reasonable alternatives.
Can a data-driven model measure incrementality?
Not from observational paths alone. It may model predictive associations; incrementality requires a suitable causal design, such as a randomized holdout.
Why do attribution reports change after a tracking update?
Updates can alter eligible events, identity matching, windows, or imports. Version the measurement contract and annotate trend breaks.
Summary
An attribution model is a documented way to distribute credit across observed touchpoints. Model choice, windows, identity resolution, and conversion definitions all shape its output. Choose a stable model for a specific descriptive purpose, make its boundaries visible, monitor data quality, and use randomized experiments or other credible causal methods when deciding whether marketing or product activity created incremental outcomes.
Sources
- Google Analytics: Attribution models
- IAB: Measurement guidelines
- UK Government: Randomised controlled trials guidance