What is Multi-touch attribution?

Multi-touch attribution is the family of approaches that distribute a conversion's credit across several touchpoints rather than assigning it all to one.

The motivation is obvious once a journey is written out. Somebody watches a video, reads a review, searches a category term, sees a retargeting ad, searches the brand, clicks, and buys. Six touchpoints, one conversion, and a single-touch rule that hands everything to one of them and nothing to the other five.

Spreading the credit is the right instinct. The difficulty is entirely in the execution, and almost none of it is in the model.

Linear. Every touchpoint gets an equal share. Simple, defensible, and wrong in an obvious way: a passing impression and a deliberate click that arrived at a checkout are not equivalent.

Time decay. Later touchpoints get more, on the reasoning that proximity to the purchase indicates influence. It is a softened last-click and carries a softened version of the same bias.

Position-based. Fixed shares to the first and last touches, with the remainder divided among the middle. Popular because it acknowledges both demand creation and demand capture, and arbitrary because the shares are chosen rather than derived.

Data-driven. The distribution is learned from journeys that converted and journeys that did not, rather than assumed in advance. This is the only member of the family that looks at the data before deciding.

The first three can be implemented in an afternoon from a clean dataset. That is not the hard part.

Every one of these models assumes the journey is visible. In practice it is fragmented.

Devices. A person browses on a phone and buys on a laptop. Without a signed-in identifier those are two people.

Browsers and apps. In-app browsing, private windows, and tracking prevention all break continuity between sessions.

Platforms. Each walled platform reports its own touchpoints and does not expose the underlying journey, so an external model cannot see what happened inside them.

Consent. Visitors who decline tracking are absent entirely, and they are not a random sample.

Offline. Phone calls, shops, conversations and anything else the business does away from a browser.

The consequence is that a multi-touch model built on incomplete journeys distributes credit across the fragment it can see and reports the result as though it were the whole. It is not less biased than last-click; it is differently biased, and less legible about where.

This is why the honest question before building one is not which model but whether the identity resolution exists. Businesses with signed-in users and a durable customer identifier have a real chance. Businesses relying on anonymous browsing across several platforms largely do not, however good the model is.

Accounts that go down this road find the work is not where they expected.

Collection. Every touchpoint has to be captured, tagged consistently and stored — which means tagging discipline across campaigns, channels and teams, maintained indefinitely.

Identity. The join across devices and sessions, which is the expensive part and the part that privacy changes keep making harder.

Reconciliation. The model's output has to be squared with the order table and with each platform's own figures, which will never agree.

Maintenance. New channels, new campaigns, changed URLs, a site migration — each breaks something, quietly.

Governance. Somebody has to own the model, defend its outputs, and explain why its numbers differ from the platforms' when a budget argument starts.

The failure mode is not a model that produces wrong answers. It is a model that produces answers nobody trusts, because when the numbers disagree with a platform's — as they will — there is no way to establish which is right, and the discussion reverts to whichever figure the arguing party prefers.

None of this measures cause.

A multi-touch model observes sequences and distributes credit across them. A touchpoint that appears in many converting journeys may be causing conversions or may be a place that people who were going to convert happen to pass through. The data cannot tell the difference, because it does not contain what would have happened otherwise.

Sophistication does not resolve this. A carefully built model on complete data is a better convention for distributing credit and still not a measurement of effect.

Most accounts are better served by three coarser instruments than by one elegant model.

A platform data-driven view for allocating between broadly comparable things, accepting its bias as roughly constant.

The blended ratio — total revenue over total spend — as the control, because it starts from the bank and cannot be double-counted.

Periodic incrementality tests on the questions that decide real money, because they measure cause and nothing else does.

That arrangement needs no identity resolution, survives every privacy change, and answers the questions that matter. It is less satisfying than a unified view of the customer journey, and it has the advantage of being obtainable.

Where multi-touch works bottom-up from individual journeys, media mix modelling works top-down from aggregates: spend by channel by period against outcomes by period, with seasonality and external factors controlled for.

The trade is almost exactly inverted. Modelling needs no identity resolution and no tracking at all, which makes it immune to every privacy change that breaks the journey-level approach. It handles offline and unmeasurable channels, because it only needs spend and results. And it is coarse — it describes channels over months rather than campaigns over days — and it needs years of history with genuine variation in spend to fit anything reliable.

For a large advertiser with long history and substantial offline activity, modelling is frequently the better investment. For a smaller one, the history does not exist and the variation was never deliberate enough to fit against.

Neither approach measures cause. Modelling infers association from aggregates; multi-touch infers it from sequences. The test remains the only instrument that withholds something and observes what changes.

Before committing to build anything, three questions settle it.

Can journeys actually be joined? Signed-in users and a durable identifier, or not.

Will the output be trusted when it disagrees with a platform? Somebody has to own that argument, in advance.

Would the same effort spent on incrementality testing answer more of the questions? For most accounts, yes.