A marketing attribution model is the rule that decides how conversion credit is split across a customer’s touchpoints. And here’s the truth the industry rarely says out loud: choosing a model is choosing what your reports will exaggerate. Every model distorts something: the job isn’t finding the undistorted one (it doesn’t exist), it’s picking the distortion you can live with.
We will elaborate on: single-touch, multi-touch rules, algorithmic models, and the newer causal family, then ends with a decision guide by business situation.
What an attribution model is (and what it isn’t)?
An attribution model is the credit-splitting rule, separate from the tracking data it runs on. Feed the same journeys to two different models and you’ll get two different “best channels.” That’s not a malfunction: it’s what a model is. Each one encodes an opinion about where influence lives in a journey. In other words an attribution is the system of beliefs you have around what caused or influenced your customers to buy your product.
What a model is not: ground truth, and not causation. A model saying a channel “drove” a sale means the rule assigned it credit, not that the sale wouldn’t have happened otherwise. The most useful mental frame: an attribution model is an accounting policy for conversions. Companies argue about accounting policies precisely because they change reported reality without changing reality.
For the wider foundations, start with what marketing attribution is.
Single-touch models: first click and last click
Last click assigns everything to the final touchpoint before conversion. It’s been the default nearly everywhere for years because it’s simple, definitive, and closest to the money. Its distortion: systematic flattery of closers: brand search, retargeting, cart emails, while the channels that created the demand get nothing. For brand-heavy businesses this is quietly expensive.
First click is the mirror: everything to the touch that opened the journey. It over-credits discovery, ignores everything that nurtured and closed, and carries a practical fragility: the true first touch often predates your tracking entirely.
Honest uses exist for both: last click for bottom-funnel audits and platform hygiene, first click for comparing demand-creation channels against each other. As a company-wide truth, neither survives contact with a multi-touch world.
Multi-touch models: linear, time decay, position-based
Linear splits credit evenly across every touchpoint. Democratic and easy to explain; its distortion is treating a two-second glance and the demo that sealed the deal as equals.
Time decay weights touches by recency: the closer to conversion, the more credit. It’s the sanest hand-set default for short sales cycles, and its distortion is gentler than last click’s but points the same way: recency isn’t the same thing as influence.
Position-based (U-shaped and W-shaped) puts heavy credit on the first and last touches (the W adds a mid-funnel milestone like a form fill), lighter credit in between. It’s an explicit opinion about journey shape, baked into a formula. If your journeys genuinely have decisive openings and closings, it fits; if not, it manufactures them.
All three see the journey, which beats single-touch. All three are still opinions. The full mechanics of the rules-based family are in first, last, linear and time decay compared. The reality around the deterministic (pre-defined) multi-touch attribution modeling is that it never gain as much traction. Although some companies adopted similar approach or even built their own multi-touch models, it never got big adoption.
Algorithmic models: data-driven, Shapley and Markov
Three approaches dominate the era of algorithmic attribution modeling. Academia initiated a different approach with Markov chain and then the industry followed with major publications which shaped the industry.
Data-driven attribution (DDA) compares journeys that converted with similar journeys that didn’t, and learns how much each touch type shifted the odds of conversion. If journeys containing a product-review visit convert far more often than matched journeys without one, the review page earns credit accordingly. This is what GA4 and Google Ads use by default: our full guide is data-driven attribution explained.
Shapley-value attribution borrows from cooperative game theory (the Shapley value, named for economist Lloyd Shapley). It asks: across every observed combination of channels, what does each channel add when it’s present versus absent? The credit split that answers this is provably the fairest division of the observed cooperation: an elegant idea, computationally heavy, and popular in home-built attribution.
Markov-chain models take the removal route: model journeys as chains of states, then ask what happens to total conversions if a channel is deleted from the chains. The bigger the drop, the bigger that channel’s credit: the “removal effect.”
All three genuinely improve on rules. All three share two honest limits: they’re black boxes to varying degrees, and they only learn from the journeys their platform can see. A model can be brilliantly fair about incomplete evidence and still mislead you.
The causal family: measuring influence, not journeys
The newest family is defined by what it refuses to do: reconstruct individual journeys at all. Causal methods estimate each channel’s influence on outcomes directly from aggregate data: spend, activity and results over time. Incrementality is the design goal, not an afterthought, and no user-level data enters the process.
This is MarSci’s home ground. Our Causal-Driven Attribution (CDA) estimates channel influence using causal inference on aggregate data. The method is peer-reviewed and open sourced. Because nothing user-level is involved, it’s untouched by cookie deprecation, works where journeys are heavily offline, and clears compliance in regulated industries where tracking-based methods can’t operate.
When tracking is broken, regulated, or simply doesn’t cover where your money goes, this family isn’t an alternative: it’s the only honest option. See MarSci’s data-driven attribution solutions.
Choosing: a decision guide by situation
- Short-cycle digital retail: data-driven attribution where your conversion volume supports it; time decay as the hand-set fallback below that threshold. Volume is the deciding factor, not preference.
- B2B long cycles: journeys run for months and close in a CRM, so tie attribution to CRM stages (position-based mapped to pipeline milestones) or go causal. Pure click-trail models mislead most exactly where deals take longest.
- Heavy offline or walled-garden spend: causal methods paired with MMM. Click trails can’t referee spend they can’t see, however clever the algorithm on top.
- Regulated industries (finance, pharma, gambling): causal, full stop. It’s the family that never needed the data you’re not allowed to use.
- Small team starting out: any conscious choice beats the unexamined last-click default. Pick one, apply it consistently, revisit quarterly as volume grows.
One switching rule applies to everyone: when you change models, restate your history under the new model. Otherwise every trend line has a seam in it where the accounting policy changed, and comparisons across the seam are fiction.
Using any model honestly
Three habits separate teams that use attribution well from teams that are used by it.
Know what your model exaggerates: every model on this page has a named distortion, and yours is working on your reports right now. Cross-check against an aggregate view before moving serious budget: a marketing mix model reads the whole mix, including what no journey tool sees, and disagreement between the views is signal, not noise. And let incrementality tests referee: a controlled experiment (geo holdout, budget test) is the only arbiter that outranks every model in the room.
The best model is the one that improves your decisions, known distortions and all. MarSci runs causal attribution and MMM as one cross-checking stack, so no single model’s opinion goes unexamined: talk to MarSci.
Frequently Asked Questions
Four families: single-touch (first and last click), multi-touch rules (linear, time decay, position-based), algorithmic (data-driven, Shapley value, Markov chains), and causal methods that estimate channel influence from aggregate data without reconstructing journeys at all. Each family trades simplicity for honesty in its own way.
Honestly, none is “accurate”: each distorts differently. Algorithmic models beat hand-set rules within tracked journeys; causal methods measure incremental influence where tracking fails. Match the model to your journey shape and data reality, and cross-check big decisions against an aggregate view.
A game-theory method (after economist Lloyd Shapley) that splits conversion credit by what each channel adds across every observed combination of channels: the provably fairest division of observed cooperation, though still limited to the journeys your data can see. Elegant mathematics, honestly bounded by the data’s coverage.
Any model that splits a conversion’s credit into fractions across touchpoints instead of giving one touch everything: linear, time decay, position-based and data-driven models are all fractional. The fractions sum to one conversion; how they’re divided is the model’s opinion.
Use one primary model consistently for decisions, keep a second view for contrast, and when you switch, restate history under the new model so trends stay comparable. Consistency is what turns a model from a report into a learning system.
