Multi-touch attribution: which model fits reality

Attribution is where a lot of marketing measurement quietly goes wrong, and it goes wrong in a specific way. Teams treat the model as if it were a measurement of truth, argue about which one is correct, and then lose confidence in all of them when the numbers refuse to agree.

The disagreement is not a sign that the models are broken. It is a sign that attribution was never a measurement in the first place. It is a set of assumptions about how to divide credit, and different assumptions give different answers because that is what assumptions do.

Multi-touch attribution splits credit for a conversion across the several touchpoints that led to it, rather than giving it all to the last click. No model is objectively correct, because attribution is a simplifying assumption, not a measurement, so the right one is the model that fits the decision you are trying to make.

Choose the one that reflects how your buyers actually decide, treat the numbers as directional, and fix the first-party data underneath before trusting any of them. Here is how to do all three.

What an attribution model is actually for

An attribution model exists to answer one practical question: given limited budget and effort, where should the next unit go. It is a decision aid, not a historical record.

That framing matters because it tells you how good a model needs to be. It does not need to reconstruct the exact causal truth of every purchase, which is impossible. It needs to be good enough to point your next allocation in a better direction than guessing.

Once you hold it to that standard rather than to an impossible standard of accuracy, the whole subject gets calmer and more useful.

The trouble starts when attribution is asked to settle credit rather than guide decisions, because credit is political. If a model's output decides whose budget grows and whose shrinks, every team has a stake in which model is chosen.

This is why attribution arguments are so often unresolvable: they are not really about the models. Keeping attribution pointed at the forward-looking question, where should the next pound go, is the single most important discipline.

Why there is no perfect attribution model

Every attribution model makes a choice about how to divide credit among the touches in a journey, and every choice distorts something. A single number cannot faithfully represent a messy, multi-step, partly-invisible human decision.

There is no fact of the matter about how much a given touch "really" contributed; the contribution is entangled and counterfactual and cannot be cleanly recovered.

What each model offers is a different lens, and each lens is honest about some journeys and misleading about others. The mature use of attribution is to hold this lightly: to know what your chosen model over-credits and under-credits, and to read its output with that distortion in mind.

The search for the model that finally tells the truth is where a lot of measurement effort goes to die. That model does not exist.

The main attribution models and what each distorts

It helps to name the common models and be blunt about what each one gets wrong.

Last click gives everything to the final touch and nothing to what created the demand. It is simple and it systematically starves the top of the funnel.

First click does the reverse, crediting the touch that started the journey and ignoring everything that closed it. Linear splits credit evenly across all touches, which treats a throwaway interaction as equal to a decisive one.

Time-decay weights recent touches more heavily, which flatters bottom-of-funnel activity and undervalues the awareness work that began the journey months earlier. Position-based rewards the first and last touch and squeezes the middle.

None of these is right, because each is a rule for dividing credit rather than a recovery of what happened. The practical way to read any of them is to ask what it would tell you to do more of and less of, and whether that instruction matches what you know about your buyers from outside the data.

Choosing the attribution model that fits your decision

Since no model is correct, the selection criterion is fit: which model's distortions are least harmful for the decision you make and the way your buyers actually move.

Start from the buying journey. If your sales are genuinely short and single-session, last click distorts little, and its simplicity is a virtue; a more elaborate model would add complexity without adding insight.

If your journey is long, considered, and multi-touch, which is typical for the deals a marketing leader cares most about, last click is actively dangerous. It systematically starves the early-funnel work that created the demand it hands full credit to the final step for.

For those longer journeys, a multi-touch model, time-decay or position-based, will make better allocation decisions. Not because it is accurate, but because its distortions punish the right behaviour less.

When a model tells you to defund the thing you know starts the journey, that is the model's distortion talking. It is a reason to change the lens, not to defund the channel. Pick the model whose blind spots you can live with, and be explicit about what they are.

The data problem under every attribution model

Underneath the choice of model sits a problem that matters more than the choice, and it is where most attribution quietly fails: the data feeding any model is incomplete before the model runs.

Tracking is lost to privacy controls and blocked scripts. Identities are not linked across the devices and sessions a single person uses. Offline touches, a conversation, an event, a referral, never enter the system at all. Demand created in places you cannot see shows up as if it came from nowhere.

Every model is therefore dividing credit over a partial and skewed picture, and a more sophisticated model applied to worse data is still wrong, just more confidently.

This is why investment in the first-party data foundation, clean event capture, identity resolution across sessions, deliberate ways to capture off-platform touches, returns more than investment in a cleverer model.

That foundation is what everything else in a marketing system depends on, which is why attribution is really an operations problem, connected to revenue operations and to whether sales and marketing share one view of the data. Fix the picture the model sees before you argue about how the model divides it.

The takeaway

Multi-touch attribution has no objectively correct model, because attribution is a simplifying assumption rather than a measurement. Keep it pointed at the forward-looking question of where the next unit of budget should go, choose the model whose distortions are least harmful given how your buyers decide, and read its numbers as directional.

Above all, fix the first-party data underneath before trusting any model at all, because a better model on a worse picture is still wrong. Attribution is a decision aid, and a good decision aid honestly held beats a perfect one that does not exist.

If your attribution numbers never quite add up and no model seems to settle it, the problem is usually the system and the data underneath rather than the model on top, which is the work of an AI marketing systems engagement.