What is Data-driven attribution?

Data-driven attribution distributes credit across the touchpoints before a conversion according to patterns a platform learned from its own data.

Instead of applying a fixed rule — everything to the last touch, everything to the first, an even split — it compares journeys that converted against journeys that did not, and assigns more credit to the interactions that distinguish them. A touchpoint appearing disproportionately in converting paths gets more; one appearing equally in both gets less.

The implementations are proprietary and differ between platforms. What the category shares is the shape: learned association rather than an assumption chosen in advance.

The fixed rules all share a defect: they encode a belief about where value sits, and that belief is wrong in a way that does not vary with the business.

Last-click hands everything to the bottom of the funnel. First-click hands everything to the top. A linear split treats a passing impression and a deliberate click as equivalent. None of them has looked at the data.

A data-driven model does look, and the result is usually a redistribution away from the final touch toward earlier interactions. For most businesses that is closer to how buying works, and it is the reason Google Ads moved to it as its default.

The practical effect on an account is that upper-funnel activity stops reporting as worthless, which changes budget decisions that a last-click view was quietly making.

It is still attribution. The model distributes credit across sequences that were observed. It has no access to the counterfactual — what would have happened had the advertising not run — because that was never recorded. A touchpoint strongly associated with conversion may be causing conversions or may simply be where people who were going to convert happen to pass.

It is bounded by one platform's view. A platform's model sees the touchpoints on its own surfaces and infers the rest. Journeys crossing several platforms are partly invisible to each of them, and each distributes the credit it can see across the touchpoints it owns.

It cannot be audited. This is the sharpest cost. A last-click figure can be recomputed from a log by somebody who wants to check it. A data-driven figure cannot: the model is proprietary, the inputs are not fully exposed, and when the number changes there is no way to tell whether the advertising moved or the model did.

It degrades with coverage. Consent requirements and tracking restrictions mean fewer journeys are observed end to end. The model fills the gaps with estimates, so an increasing share of what it distributes is modelled rather than seen.

Keep a deterministic view alongside. Running last-click in parallel gives a figure that can be checked, and the gap between the two is itself informative. A widening gap means the model is crediting earlier touchpoints more heavily than before, which is a measurement change and should not be read as upper-funnel activity improving.

Note when it changes. Platforms update these models. A step in the reported figures with no corresponding change in delivery is a model update, and comparing across it is comparing two instruments.

Do not sum across platforms. Each platform's model credits the touchpoints it can see, and two platforms modelling the same journey will both claim parts of it. The blended ratio remains the control.

Watch the volume behind it. A model learned from a handful of converting journeys is a model learned from a handful of converting journeys, and an account near that end is getting figures that will be noisy in ways that look like performance.

The honest summary is that a data-driven model is the best available convention for distributing credit, and distributing credit is not the same as measuring effect.

For allocation between broadly comparable things — which campaign, which creative, which audience — it is the right instrument, because the bias is more even than a fixed rule's and the decisions are small and reversible.

For the questions that decide real money — whether a channel should exist, whether retargeting adds anything, whether brand spend pays — it is the wrong instrument however sophisticated it is, because those are causal questions and no attribution model can answer one.

The working arrangement is the same as it is for every part of this discipline. Steer with attributed figures, control with the blended ratio, and settle the expensive questions with a test that withholds the advertising and measures the difference.

The move from a fixed rule is worth anticipating, because several of its effects look like performance changes and are not.

Conversion counts move between campaigns. Totals stay roughly stable while the distribution shifts away from branded search and retargeting toward prospecting and upper-funnel activity. Nothing about delivery changed.

Fractional conversions appear. Credit is distributed rather than assigned, so a campaign can report a conversion count with a decimal in it. This confuses people the first time and is correct.

Automated bidding retargets. A conversion-based strategy optimises against the conversions it is told about, so redistributed credit changes what the model is pursuing, and the campaign enters a period of unstable delivery while it adjusts.

Historical comparisons break. A report drawn across the switch is comparing two instruments.

The sensible handling is to treat it as a material change: make it deliberately, note the date, hold budget steady through the following weeks, and do not read the transition period as a result.

Name the model and the window on every figure that leaves the team.

Keep a deterministic view in parallel and watch the gap rather than either number alone.

Record platform model updates where they are announced, so a step in the data has an explanation attached.

Never sum across platforms.

Schedule a test rather than waiting for an argument that needs one.

Last-click gives every conversion to the final ad interaction before it; a data-driven model distributes the credit across the touchpoints using patterns learned from journeys that converted and journeys that did not. The case for the model is that a fixed rule encodes a belief about where value sits before looking at anything, and that belief is wrong in a way that does not vary with the business. A data-driven model looks, and the redistribution is usually away from the final touch toward earlier interactions — closer to how buying actually works, and the reason Google Ads moved to it as its default. Two things keep the comparison honest. It is still attribution: it distributes credit across observed sequences and cannot see what would have happened without the advertising, so it inherits every limitation of the category. And it cannot be audited — a last-click number can be recomputed from a log, and this one has to be taken on trust, which means that when it moves there is no way to tell whether the advertising changed or the model did.