A purchase rarely has one cause. A buyer sees a video advertisement, searches a week later and clicks an organic result, returns through a retargeting ad, opens an email and buys from a brand search. Five touches, one sale: attribution is the rule that divides the credit. Google Analytics describes it as assigning credit for important user actions to the ads, clicks and factors along the path, and in its current version offers three rules: data-driven, which uses a model trained on converting and non-converting paths; paid and organic last click, which gives the whole credit to the final channel before the action and ignores direct visits unless nothing else is in the path; and Google paid channels last click, which gives it to the last Google Ads touch.
The rule chosen changes the dashboard without changing a single sale, which is why disputes about attribution are in substance disputes about budget. Under last click the brand-search campaign looks like the best investment the business makes, because people who have already decided search the name on the way in; under a data-driven model the video and the content that started the path receive some of that credit and brand search shrinks. Neither view is the truth. The model reports a share of credit, and credit is a convention; what the business wants to know is incrementality, how many of those sales would not have happened without the channel, and no division of touches answers that.
| Model | Rule | Which channels it favours |
|---|---|---|
| Last click | All credit to the final touch | Brand search, retargeting, email to existing customers |
| First click | All credit to the first touch | Awareness channels, content, display |
| Linear, position-based, time decay | Credit divided by a fixed shape across the path | None in particular; the shape is arbitrary |
| Data-driven | Credit divided by a model the platform trains and does not publish | The platform's own channels, by construction |
| Holdout or lift test | Withhold the channel from a random group, compare | The causal answer, at the cost of a test |
Three habits make attribution useful despite its limits. Read the same report under two models and treat the channels whose credit swings most as the ones whose value is least understood. Run a holdout where it matters, which is an A/B test at the level of a channel rather than a page. And for spend that cannot be tracked by click at all, television, podcasts, out-of-home, use marketing mix modelling, a regression of sales on spend over time, which has returned to fashion as click-level tracking has shrunk. The analytics and attribution courses cover the models; the Google Analytics certification and the Ads Measurement one cover the settings and the conversion tracking the models run on.
What attribution cannot see
A recommendation at dinner. A review read on a phone that was never logged in. A purchase made in a shop after a search on a laptop. Consent-declined visits. Every model divides credit among the touches it recorded, and the recorded touches are a shrinking share of the path.