What Marketing Attribution Actually Is

Most marketing teams use the word attribution as if it described a single, knowable fact about where their sales come from. It does not. Attribution is a bookkeeping decision. It takes the touchpoints a buyer passed through before converting, an ad they clicked, an email they opened, an organic search they ran, and divides credit among them according to a rule somebody chose in advance. Change the rule and the credit moves. Nothing about the underlying customer behavior changed at all.

The problem attribution is trying to solve

A typical purchase is not a straight line. Someone hears about a product from a podcast, searches the brand a week later, clicks a retargeting ad, ignores it, comes back through a Google search, and finally converts after a discount email lands. Six channels touched that one sale. The finance team wants to know which channels to fund next quarter, and the honest answer is that all six played some role that is genuinely hard to untangle. Attribution exists to give that messy reality a number, so budgets can be argued over with something other than gut feel.

The catch is that assigning a number requires a theory of how influence works, and there is no neutral theory available. Crediting the first touch assumes discovery is what matters. Crediting the last touch assumes the closer is what matters. Spreading credit evenly assumes every step counts the same. Each of those is a guess dressed up as a calculation.

The single-touch models

The two oldest models assign all the credit to one interaction. They are simple to run and simple to misread.

• First-touch: gives the entire conversion to the very first interaction the buyer had, which inflates whatever sits at the top of the funnel like brand awareness or content
• Last-touch: gives the entire conversion to the final interaction before the sale, which inflates bottom-of-funnel channels like branded search and retargeting that often just catch demand other channels created

Both are still widely used because they are easy to explain to a board, and last-touch in particular remains a default in most analytics tools. Their weakness is the same in mirror image. A first-touch report will tell you to pour money into awareness and quietly starve the channels that actually close. A last-touch report will tell you the opposite. Neither is lying. They are answering different questions and presenting the answer as if it were the only question.

The multi-touch models

Multi-touch attribution tries to fix the obvious flaw by spreading credit across several touchpoints instead of dumping it all on one. The common variants are still rules, just more elaborate ones.

• Linear: splits credit equally across every touch in the path, which treats a forgettable banner impression the same as the demo that sealed the deal
• Time-decay: weights touches closer to the conversion more heavily, on the assumption recent contact mattered more
• Position-based: front-loads and back-loads the path, often giving 40% to the first touch, 40% to the last, and dividing the remaining 20% among the middle
• Data-driven: uses a model trained on a company's own conversion paths to estimate each touch's contribution, rather than a fixed weighting a human picked

Data-driven attribution is the most defensible of the bunch because the weights come from patterns in the data rather than a hunch. It is also the hardest to audit, since the logic lives inside a model most marketers cannot inspect. In 2023 Google retired its first-click, linear, time-decay, and position-based models in Ads and Analytics, citing adoption below 3% of conversions, and pushed everyone toward data-driven attribution while keeping last-click available. That move tells you something. Even the company that built the rules-based models decided most of them were not worth maintaining.

Why the models never fully agree

Run the same quarter of data through five attribution models and you will get five different rankings of your channels. This is not a bug to be configured away. It is the direct result of the fact that each model encodes a different belief about influence, and the data alone cannot settle which belief is correct. The journey records that a touch happened. It does not record whether the sale would have happened anyway without that touch, and that counterfactual is the thing budget decisions actually depend on.

This is the core limitation worth internalizing. Attribution measures correlation along a path. It shows which channels tend to appear before conversions. It cannot show which channels caused conversions, because causation requires knowing what would have happened in their absence, and a log of clicks contains no such information. A brand-search ad that catches buyers who were already going to type the brand name will look like a star in almost every model, right up until someone turns it off and notices the sales barely move.

What broke the data underneath

Attribution also got harder for reasons that have nothing to do with the models themselves. Apple's App Tracking Transparency, introduced with iOS 14.5 in April 2021, let users decline cross-app tracking, and most did, which cut off a large share of the signal that user-level mobile attribution depended on. Walled gardens like the major social and search platforms report their own conversions inside their own systems and do not share the raw paths, so a marketer stitching together a journey is often working with fragments. It is worth noting that one widely predicted disruption did not arrive. Google decided in 2025 to keep third-party cookies in Chrome rather than deprecate them, so the cookie itself is still around even as the rest of the tracking picture has fragmented.

The practical effect is that the inputs to attribution are noisier than the confident dashboards suggest. Gaps get filled with modeling and assumptions, and those assumptions compound with the assumptions already baked into the model. The output still looks precise. It is worth treating that precision with suspicion.

How to actually use it

None of this means attribution is useless. It is a reasonable tool for spotting directional patterns, comparing two campaigns run under the same model, and noticing when a channel that used to appear in conversion paths suddenly stops. The mistake is treating the credit split as ground truth and reallocating real money on the strength of it. Attribution can tell a team where to look. It cannot tell them what is causing the result, and confusing the two is how budgets end up funding channels that take credit without doing the work. The honest version of the report comes with a caveat the dashboard never prints, which is that this is one model's opinion, and a different model would have a different one.