Attribution models, explained without the jargon

The question is: this conversion was preceded by several interactions, so which of them gets the credit, and how much.

That is the whole job. The model does not decide whether the conversion is recorded, how many conversions one click may produce, which date the conversion appears on, or whether an interaction on a non-Google channel is eligible for credit at all. Four different settings decide those, and every one of them gets called "attribution" in meetings.

So when someone says the attribution is wrong, the first job is to find out which of the five things they mean. In my experience it is almost never the model. It is usually the window, the counting rule, or the fact that two systems are looking at different sets of interactions.

Google keeps the definitions in one place. Its page on attribution models covers the Ads side, and Analytics documents its own attribution concepts separately, which is itself a hint that the two products are not describing the same thing.

The window is the one most often mistaken for the model. Google defines it precisely: a conversion window is the period after an ad interaction during which a conversion is recorded in Google Ads. Outside that period the conversion is not credited late, it is not credited at all, so a window shorter than your sales cycle deletes your slowest buyers from the data before any model gets to see them.

The counting option is the one that produces the strangest arguments, because it changes the numerator rather than the split. Google offers Every and One, and the default differs depending on how the conversion action was created. A form on Every counts three submissions from one determined visitor as three conversions, and no attribution model reconciles that with a CRM holding one record.

The last row is the one nobody sets and everybody argues about. Google Ads can only credit interactions with Google ads. It has no mechanism to credit an organic visit or an email click, because those are not things it observes. That is not a modelling decision, it is the boundary of the product.

Google's own note on the deprecated four is blunt: first click, linear, time decay and position-based models are no longer supported, and conversion actions using them were upgraded to data-driven attribution, with last click available as the alternative.

This matters if you maintain a reporting document written before that change. A monthly deck that still explains a position-based model is describing a setting the account cannot have, and someone will eventually check.

It matters more for year-over-year comparison. A period reported under a model that no longer exists is not directly comparable to a period reported under the current one, and nothing in the interface warns you when you drag the date range back past the switch.

The row I would draw a box around is the middle one. Google describes data-driven attribution as comparing the paths of customers who convert against the paths of customers who do not, and says each data-driven model is specific to each advertiser. Two consequences follow that are worth stating out loud. Your model is not the same object as your competitor's model, so borrowed benchmarks are meaningless. And the split it will produce for your account is not published in advance, so anybody who tells you what percentage of credit a channel will receive is guessing.

Take one real path. A visitor clicks a generic search ad on the first, clicks a display ad on the ninth, searches your brand name and clicks that ad on the fourteenth, then buys.

Under last click, the brand ad receives 1.0 conversions and the other two receive 0.0. The report will then show the brand campaign with an excellent cost per acquisition and the generic campaign with none, which is exactly the arithmetic that gets a prospecting campaign cut.

Under data-driven, credit is divided across those interactions using your account's data, and the fractions are computed by Google rather than chosen by you. That is the honest description. I cannot tell you what the fractions will be, and neither can anybody else outside Google, because the calculation runs on your data and is not published as a formula you can apply by hand.

What you can do is measure the change. Record cost per acquisition per campaign for a full period before switching, then again after, and treat the difference as a reallocation of the same total rather than as performance. It is the same conversions, described differently. If the total moves a lot, something other than the model moved too.

Google Ads and Analytics are not two implementations of one idea. They answer different questions, and the settings are configured in two places: the attribution settings in Analytics are selected separately from the model on a Google Ads conversion action, so a change in one quietly widens the gap in the other.

The scope difference is the big one. Analytics can credit organic search, direct, email and referral traffic. Google Ads cannot, because it does not observe them. Any conversion Analytics credits to a non-paid channel simply will not appear in the Ads number, and that is designed rather than broken.

The unit difference is the second one. The two products do not count the same object in the same way, and no configuration makes them identical. That is the core of why the two platforms rarely agree on a conversion total, and it is worth reading before anyone opens a ticket.

So the practical rule is: never compare a model in one product to a model in the other and conclude anything. Compare a product to itself over time, and compare products only on the size and stability of the gap between them.

Attribution divides credit for conversions that were recorded. It does not tell you whether an interaction caused the conversion. Those are different claims and the data cannot distinguish them.

A brand search ad credited with a conversion under last click may have been the last thing a decided buyer touched. Under any model, it is credited. Nothing in the report distinguishes an ad that persuaded someone from an ad that was merely present at the end.

That limit is not a flaw in the model, it is what attribution is. The only way to ask the causal question is to change the exposure for one group and not another and compare the outcomes, which is a separate exercise with its own cost and its own minimum size.

So when a client asks what would happen if we turned this campaign off, the honest answer is that the attribution report does not know. Say that, then propose the test, rather than quoting a credited conversion count as though it were an answer.

First sentence: every model is looking at the same conversions, and changing the model changes who gets credited, not how many there were.

Second sentence: the total is the number to hold anyone accountable for, and the per-campaign split is a lens for deciding where to spend next.

That framing survives contact with a finance team, because it separates the thing that ties to money from the thing that guides allocation. Presenting a model change as a performance improvement does not survive, and it costs you the room the first time somebody notices the total did not move.

Add one more line if the account has just been switched: expect campaign-level cost per acquisition to move in both directions this month, and expect the account total to stay roughly where it was. Predicting that in advance is worth more than explaining it afterwards.

Three signs the model is a distraction. The conversion action counts an action nobody can defend as a business outcome. The window was inherited rather than chosen. Or two systems disagree by a factor that changes every month.

In a multi-location business there is a fourth, and it outranks the others: conversion events that fire without a location parameter make per-site cost per acquisition unmeasurable no matter which model is selected. Fixing the model there is polishing a number you cannot break down.

That was the shape of it at a clinic group with more than fifteen sites, where the fix was at the event layer and the payoff was per-location reporting that runs from live data rather than a better model.

And when four systems disagree and nobody can say which is supposed to be right, that is not a model question either. It is a conversion attribution audit: quantify each gap, separate the structural variance from the defects, and name one authoritative source per decision. Changing the model underneath an unreconciled definition set just moves the argument somewhere new.

One last check before you touch anything. Read the conversion window on every conversion action and write down what it is, because that setting governs whether the data reaching your model is complete, which is the same lag question that governs every other read of the account.