What is attribution modeling

Guide · Glossary & Definitions · 4 min read · last verified 2026-07-27

Attribution modeling is the set of rules a team uses to decide which marketing touchpoints get credit for a conversion. Note the verb: decide, not discover. A buyer's path to purchase happened however it happened; the model does not reveal what caused what — it applies a policy for splitting credit among the interactions that got recorded. Understanding attribution starts with accepting that every model is an accounting choice made before the data arrives, and none of them is the truth.

Every model is a policy, not a measurement

This framing matters because attribution outputs look like measurements. They arrive as tidy numbers next to channel names, with the visual authority of a thermometer reading. But change the model and the same underlying paths produce different numbers, while nothing in the world has changed. A measurement that moves when you change your opinion about credit is not a measurement of the world; it is a report of your policy applied to the world.

That does not make attribution useless. Companies run on accounting choices — depreciation schedules, cost allocations — and run better for having them. The failure mode is forgetting the choice was a choice, and defending a credit policy as if it were a causal finding.

The model families

Each common family answers the credit question with a different rule, and each rule has a personality:

• Last touch gives everything to the final recorded interaction. It flatters whatever sits closest to purchase and erases everything that created the intent.
• First touch gives everything to the earliest recorded interaction. It flatters discovery and erases everything that carried the deal home.
• Linear spreads credit evenly across recorded touches. It offends no channel and informs no decision sharply.
• Time decay weights later touches more heavily. It encodes the assumption that recency implies influence — an assumption, not a finding.
• Position-based concentrates credit at the first and last touch. It encodes the opposite guess: that openings and closings matter most.
• Data-driven or algorithmic models learn credit weights from patterns in observed paths. These are often marketed as objective. They are better described as a policy the machine wrote: they inherit every gap in the recorded data and add a layer of opacity about how the weights arose.

None of these is wrong, exactly. Each is a defensible answer to an unanswerable question — which is why the choice among them tends to reveal more about a team's politics than about its buyers.

The blindness every model shares

All attribution models allocate credit only among touches they observed. Influence that left no record — a peer recommendation, a private community thread, a mention inside an AI-generated answer — does not lose its credit; the credit gets silently redistributed to whatever was recorded, usually the click nearest the purchase. The visible channels absorb the reputation of the invisible ones.

This is the deepest reason attribution debates feel endless: the models are arguing over how to divide a pie with slices missing, and the missing slices appear to be growing as more of the buying journey moves into untrackable spaces. What is dark social maps those spaces, and Why AI referral traffic is undercounted examines the newest of them. Any attribution readout should be captioned, at least mentally: among the things we could see.

Which model should you use?

The honest reframe: which bias can you live with, and can you state it out loud? Some practical guidance follows from that.

Match the model to the decision it funds. A team optimizing closing efficiency can tolerate last touch; a team investing in demand creation will be actively misled by it. Consider running two contrasting models — one biased early, one biased late — and treating the space between them as the honest range rather than pretending either endpoint is exact. Keep whatever model you pick stable, because a policy applied consistently at least yields comparable trends, while a policy that changes quarterly yields nothing. And when reporting upward, name the model and its known blind spots in the same breath as the numbers; How to report marketing to a CFO argues that disclosed assumptions are what separate credible reporting from advocacy.

Why attribution is controversial

Because credit is budget. The model that assigns credit effectively assigns next quarter's spending, which means every channel owner has a stake in the rules — and the arguments that follow are frequently budget negotiations conducted in the vocabulary of data science. Recognizing this cools the debate considerably: the fight is rarely about which model is true, since none is, but about which policy the organization can commit to and interpret with open eyes.

The second source of controversy is the causation gap. Attribution, however sophisticated, records proximity. Whether a credited touch actually changed the outcome is a different question with a different discipline attached — What is incrementality covers the experimental complement that attribution quietly depends on for calibration.

Using attribution without being used by it

Held properly, attribution modeling is an accounting instrument: consistent, stated, useful for trends, silent on causation, and blind to the untracked. The teams that get value from it tend to pair it with direct observation of the surfaces it cannot see — asking customers how they actually heard of them, and checking presence where recommendations now happen, including inside AI answers, which is the surface Magrios measures directly rather than inferring from click residue. The model then becomes what it always was: one witness among several, useful precisely to the degree that nobody in the room mistakes it for the judge.