Attribution Models, and Which One Changes What You Do

Almost every explanation of this lists the models, describes each one, and stops. You finish it knowing the names and still not knowing what to do.

The question worth asking is narrower: would switching model change a decision you are about to make? For a lot of accounts the answer is no, and knowing that is worth more than knowing the taxonomy.

The problem attribution is trying to solve

Somebody sees a paid ad, does not click. Later they search your brand name and click. Later still they click an email and buy.

Three touchpoints, one sale, and the sale cannot be divided by any method that is objectively correct, because there is no fact of the matter about how much each interaction contributed. Attribution is a set of conventions for splitting credit, not a measurement of causation.

That sentence is the whole subject. Every model is a rule, chosen by somebody, and the differences between them are differences of convention.

The families, without naming a product's current menu

Rather than a list of the options in one platform's interface, which changes, here are the shapes.

Give all the credit to the last interaction. Simple, universally understood, and it systematically flatters whatever tends to happen last. That is usually brand search, direct visits and retargeting, which are the channels closest to a decision that was already made.

Give all the credit to the first interaction. The mirror image, and it flatters whatever tends to happen first, which is usually the channels doing discovery.

Split it evenly across every interaction. Treats a passing impression and a decisive comparison as equal, which is defensible and is nobody's actual belief.

Weight it by position or by time, giving more to the first and last, or more to whatever happened most recently. These are compromises and they are as arbitrary as the two extremes, with more arithmetic on top.

Or model it. Use observed patterns across converting and non-converting paths to distribute credit. Better in principle, opaque in practice, and the output cannot be checked by hand, which means you are trusting a method you cannot inspect.

Which of these your platform offers, and what it calls them, has changed more than once. Check your own platform's current documentation rather than any article's list, including this one, which deliberately does not give you one.

When the model genuinely matters

Three conditions, and you need all three.

Multiple channels running at once, with meaningfully different roles. If almost all of your paid activity is one channel, changing how credit is divided between channels changes very little.

Long or multi-touch paths. If most conversions happen on the first visit, there is nothing to divide.

And enough volume for the difference to be stable. A change in model applied to a small number of conversions produces a change in the numbers that is indistinguishable from noise.

If you do not have all three, the model is not your problem, and time spent on it is time not spent on the measurement issues that actually distort your numbers.

What actually distorts your numbers more

For most accounts, in order.

Broken attribution. A redirect stripping a click identifier makes paid traffic invisible regardless of which model you selected. This site covers finding that separately, and it changes the picture far more than any model choice.

Conversion actions counting the wrong thing. A conversion measuring button presses rather than completed submissions inflates every model equally.

Cross-device and consent gaps. Journeys that span devices, or visitors who decline measurement, leave paths the model cannot see. No model reconstructs data that was never collected.

And a large unattributed segment. Sessions the system could not identify go into a general bucket, and our article on what direct traffic actually is covers what ends up there and why.

Fix those before touching a model. A different rule applied to incomplete data is a different wrong answer.

The test that answers the question for your account

Ten minutes, and it is worth more than reading five model descriptions.

Take a decision you are about to make. Not a general question. A specific one: whether to keep funding a channel, whether to shift budget, whether a campaign is working.

Look at the numbers under two models, if your platform lets you compare.

Ask whether your decision changes.

If it does not, the model is not your problem and you can stop. That is the most common outcome and it is a legitimate result rather than a failure to investigate properly.

If it does, you have found something real, and the thing you have found is usually one specific channel whose apparent value depends heavily on how credit is divided. That channel is the one to look at, not the model.

Two things to hold on to

Consistency beats correctness. A number reported one way for a year, understood by everybody who reads it, is more useful than a better number that changed in March and made every comparison meaningless. If you switch, note the date and say so in the report, because the change will otherwise be read as performance.

And the model does not change what happened. The same people took the same actions and bought the same things. What changed is a bookkeeping convention. Treating a model switch as an improvement in performance is the specific error worth avoiding, and it is the one most likely to end up in a deck.

Where this leaves you

Most accounts should pick a model, understand what it flatters, and leave it alone. Last-interaction flatters what happens last; first-interaction flatters what happens first. Knowing which way your chosen model leans is most of what you need.

Then spend the effort on whether the underlying data is complete, because a model is a rule for dividing what you collected, and the collection is where the real errors live. Our article on choosing what a conversion should measure in the first place is the better use of the same hour.

Similar Posts