Attribution fundamentals: models, biases, and what attribution cannot tell you

Attribution answers one question: given a conversion and the sequence of marketing touchpoints that preceded it, how should the credit be divided? Every model — from last click to Shapley — is a different answer to that question, and none of them is “the truth.” This guide covers the two model families, the biases each one bakes in, and the boundary that matters most in practice: the line between attribution (an accounting convention) and incrementality (a causal claim).

The raw material: paths

A conversion path is the ordered list of touchpoints a user was exposed to before converting — [paid_search, email, direct] → purchase. Before any model runs, three upstream systems have already shaped the paths, and errors here dwarf model choice:

• Identity decides whose touches join into one path. Cookie-scoped paths fragment across devices and Safari’s cookie caps; person-scoped paths depend on login coverage.
• Campaign tagging decides what each touch is called — inconsistent UTMs (see the Campaign URL Standard) turn real channels into (unassigned) noise.
• Channel grouping (and the classification standard behind it) decides the granularity at which credit is assigned at all.

A perfect model over broken paths produces confident nonsense. Fix the inputs first.

Family one: rule-based (position) models

Rule-based models assign credit by position or recency with a fixed formula. They are deterministic, cheap, explainable — and each one is systematically wrong in a known direction:

Model	Rule	Systematic bias
Last interaction	100% to the final touch	Over-credits bottom-funnel, brand search, direct
Last non-direct	100% to the last non-direct touch	Hides real direct demand
First interaction	100% to the first touch	Over-credits prospecting/top-funnel
Linear	1/n to every touch	A stray impression equals a high-intent click
Time-decay	Weight halves every H days before conversion	Under-credits early discovery
Position-based (U-shaped)	40/40 to first and last, 20% spread across the middle	The 40/40/20 split is arbitrary

The exact formulas (including time-decay normalization and the position-based edge cases for 1–2 touch paths) are written out in the Attribution Models standard; every number a vendor shows you for these models is reproducible from those definitions.

The important property of this family: the data never talks back. Last click will report that brand search “drives” conversions regardless of whether it does, because the rule guarantees it. Rule-based models measure where in the path a channel tends to sit — not what it contributes.

Family two: algorithmic (data-driven) models

Algorithmic models estimate contribution from the data — crucially, from converting and non-converting paths, which rule-based models discard. Two methods dominate, and they are what vendors mean by “data-driven attribution”:

• Markov removal effect. Model journeys as a Markov chain; for each channel, delete it from the graph and recompute the conversion probability. Channels whose removal craters conversions were load-bearing and earn proportionally more credit. Order-aware, scales well, needs enough path volume for stable transition estimates.
• Shapley value. Treat channel subsets as coalitions and average each channel’s marginal contribution over all orderings — the unique allocation that is efficient, symmetric, and gives zero to channels that never add value. Order-agnostic and axiomatically fair, but combinatorial in the number of channels. GA4’s data-driven attribution and Adobe’s Algorithmic model in Attribution IQ are in this lineage.

Both algorithms, with worked math, live in the standard; the free Attribution Simulator runs the rule-based set and both algorithmic methods over the same paths so you can see how far the allocations spread.

Algorithmic models are better accounting — they are still accounting. They remain functions of observed paths, so anything the observation layer misses (blocked beacons, cross-device gaps, walled-garden view-through) is invisible to them, and they inherit every identity and tagging flaw upstream.

The settings that quietly decide the answer

Two models with the same name can disagree because of configuration, not math:

• Lookback window. A 30- versus 90-day window changes which touches exist at all.
• Conversion definition. Attribution to a purchase and attribution to a lead event over the same traffic produce different channel mixes.
• Counting scope. Whether paths are built per cookie, per person, or only across one vendor’s ad surfaces (platform-reported numbers credit only touches that platform can see — every ad platform’s dashboard sums to more than your actual conversions).

When two tools disagree, compare these settings before debating models.

Attribution is not causation

This is the honesty clause. Attribution — including data-driven attribution — is correlational. It reallocates credit among observed touches; it cannot say what would have happened without a channel. The canonical failure is retargeting: users deep in the funnel get retargeted, convert at high rates, and every observational model hands retargeting generous credit — including for conversions that would have happened anyway. The model is not broken; the question is wrong.

The causal question — what did this channel add? — is answered by incrementality testing:

• Holdout experiments — withhold the channel from a random user group, compare conversion rates.
• Geo experiments — turn spend on/off in matched regions; measure the lift.
• Marketing mix modeling (MMM) — top-down regression of outcomes on spend, useful where user-level data is unavailable, at the cost of coarse granularity.

Mature teams run both layers and reconcile: attribution allocates day-to-day, experiments calibrate it. If attribution says a channel drives 30% of conversions and a holdout shows near-zero lift, the holdout wins — and the honest response is to discount that channel’s attributed credit going forward, not to switch to a model that flatters it.

An honest operating posture

1. Match the model to path volume. With thin multi-touch data, linear or position-based is fine — the marginal insight from algorithmic models is smaller than your input noise. Adopt Markov/Shapley once tagging and identity are reliable.
2. Budget against ranges, not points. Run several models side-by-side (the Simulator exists for this). Channels that look strong under every model are safe; channels whose credit collapses under a model change deserve an experiment, not more spend.
3. Treat big model disagreements as diagnostics. A large Markov-versus-Shapley gap usually indicates a low-volume channel or a tracking gap, not a modeling philosophy problem.
4. Reserve causal language for experiments. “Attributed 500 conversions under DDA” and “incrementally drove 500 conversions” are different claims; keeping the vocabulary separate keeps stakeholder decisions honest.
5. Don’t let the model become the target. Teams optimize toward whatever the model rewards; under last click that means bidding on your own brand terms. Periodically ask whether the model is measuring the marketing or the marketing is farming the model.

Attribution done honestly is a useful ledger: it keeps spend allocation consistent, comparable, and arguable in one shared frame. It stops being useful the moment anyone forgets it is a convention — one you should choose deliberately, document, and stress-test against reality.