Here is a sentence that has moved millions in ad budget: "the data shows brand search converts best." And here is what the data actually showed: that the last thing people click before buying is often your own name. Brand search did not cause those sales; it was standing nearest the register when they happened. That is last-click attribution, it is still the default mental model in most companies, and it is lying to you in a specific, fixable way.
Let us take the models one at a time, honestly, and then talk about what they can never do.
What an attribution model actually is
A conversion happened; several marketing touches preceded it. An attribution model is a rule for distributing the conversion's credit across those touches. That is all it is: an accounting convention. Nothing in the model observes causation, tests counterfactuals, or knows whether the buyer would have purchased anyway. Keep that sentence nearby; every dashboard in this genre is quietly hoping you forget it.
The lineup, and what each one rewards
- Last-click gives everything to the final touch. It systematically flatters bottom-funnel channels: brand search, retargeting, cart-abandonment emails, everything that harvests demand. Channels that create demand (the first article read, the podcast mention) score zero. Optimizing to last-click starves your future pipeline to feed your checkout, which feels great for exactly two quarters.
- First-click gives everything to the opener: the mirror-image distortion, flattering awareness channels and ignoring everything that nursed the sale home. Rarely used alone, useful as a contrast view.
- Linear, time-decay, position-based spread credit by arithmetic: equally, or weighted toward recency, or toward the ends. More balanced, equally arbitrary. The weights encode someone's opinion, not evidence.
- Data-driven attribution (GA4's default now) uses machine learning over your conversion paths, comparing journeys with and without a given touch to assign fractional credit. Genuinely better than fixed rules, and worth using as your consistent lens. Also worth deflating: it still only sees tracked touchpoints, it works on correlations rather than experiments, and its reasoning is a black box you cannot audit (modeled numbers deserve calibrated trust). Better arithmetic on the same partial evidence is progress, not truth.
What every model is blind to
Three blindnesses, shared by all of them, from crude to clever:
- Untracked influence. Word of mouth, communities, podcasts, dark social, the recommendation over lunch. These channels create real buyers and appear in no path. Whatever credit they earned gets redistributed to whatever was trackable, silently inflating everything you can measure at the expense of everything you cannot.
- Baseline demand. The customer who was buying anyway still shows a path, and the path still gets full credit. This is how retargeting perennially looks brilliant: it stands next to people already walking through the door. Attribution cannot distinguish persuasion from presence.
- Shrinking visibility. The tracked path itself is decaying: consent, blockers, iOS all trim what any model gets to see, so the redistribution problem from point one grows every year.
What to actually do
Use one model, consistently, as a comparison lens. Data-driven where available, and the choice matters less than the consistency: a stable lens shows movement honestly even when its absolute numbers are fiction. Switching models mid-year to whichever flatters the current strategy is the analytics version of moving the goalposts, and everyone in the room knows it.
Then, for decisions with real money attached, run the test no model can fake: incrementality. Pause the channel in one market, or holdout-test the audience, and watch actual orders. If "your best channel" gets paused and revenue does not move, you have learned what three years of attribution reports could not tell you. Incrementality is slow, coarse, occasionally humiliating, and the only method on this page that measures causation instead of adjacency. The budget version: one deliberate test per quarter on your biggest or most-doubted channel beats continuous faith in any dashboard.
And keep the store's order count as the anchor (the reconciliation habit): models redistribute credit for reality; they do not define it. Attribution is a map of the roads your tracking can see, drawn by formulas with opinions. Navigate with it, argue with it, but when the map and the terrain disagree, the terrain is the one paying your invoices. If your company's channel budget currently rests on last-click alone, running the same quarter through a second model, then picking one channel for a real holdout test, is a month of mild discomfort that regularly re-prices entire marketing lines, and it is exactly the kind of honest measurement work worth doing before the next planning cycle locks numbers in.