Attribution model

An attribution model is the rule that decides which touchpoint gets the credit when somebody finally buys. Almost nobody buys on first contact, so there is a sequence of adverts, searches, posts and visits behind each sale, and the model is how a platform divides one conversion across several of them.

In short

• The rule for splitting credit across the steps before a sale.
• It changes your reports and your bidding, not your revenue.
• Google Ads now offers data-driven and last click. The rest were removed.
• Every model is a simplification, so pick one and stay on it.

Why the credit has to be divided at all

Consider an ordinary path. Somebody sees an Instagram post, does nothing. Searches your category a week later, clicks an advert, reads, leaves. Searches your business name eleven days after that and buys.

One sale, three touchpoints. Give it all to the last and the brand name search looks like an extraordinary performer while the post looks worthless, even though the post is why the person knew the name. Give it all to the first and paid search looks pointless. Neither reading is honest, and both get used to make budget decisions.

That is the whole problem attribution exists to address, and it does not solve it. It just makes the assumption explicit, which is more than most tracking setups manage.

What Google actually offers now

The list is shorter than most published advice suggests, because Google removed four models in 2023.

In Google Ads, data-driven and last click remain. First click, linear, time decay and position based are gone, and conversion actions that used them were moved onto data-driven. Data-driven is the default for most conversion actions, which means many advertisers are using it without having chosen it. Rather than applying a fixed rule, it compares the paths of people who converted against the paths of people who did not, and assigns credit according to the patterns it finds.

In GA4 the same four were removed, leaving a cross-channel data-driven model, a cross-channel last click model, and a last click model restricted to Google paid channels. The cross-channel models consider organic search, social and referral traffic as well as advertising, which is why GA4 and Google Ads can look at the same sale and describe it differently.

Words you will hear

• Touchpoint. One interaction on the path to a sale.
• Lookback window. How far back the model may look. Older touchpoints are ignored entirely.
• Conversion window. How long after a click a conversion may still be counted to it.
• Data-driven. Credit assigned by observed patterns rather than by a fixed rule.
• View-through. Credit for an advert that was seen and not clicked.
• Modelled conversions. Estimated rather than observed, filling gaps left by consent refusals and cross-device journeys.
• Direct. Traffic with no recorded source, which is where broken tracking goes to hide. Tagging with UTM parameters removes some of it.

What the model changes, and what it does not

It changes three things. The numbers in your reports. Which campaigns appear to be working, and therefore what their conversion rate looks like. And, because automated bidding optimises towards reported conversions, where the money goes next.

It changes nothing about how many people bought. This sounds obvious and is routinely forgotten, because a model change produces a visible shift in the numbers that looks exactly like a performance change. The safeguard is to note the date of any model change somewhere you will find it later, so that the step in the graph has an explanation attached to it when somebody asks in March.

The second safeguard is consistency. A business that switches model whenever the current one is unflattering has no trend data at all, only a series of incomparable periods.

Living with it honestly

For most small businesses the correct amount of attribution work is small, and the risk is not choosing badly. It is believing the output completely.

Watch for the trap that last click sets. Under it, channels that introduce people to a business look weak and channels that catch people already looking for it look strong, so the reasonable-seeming decision is to move everything into the second kind. That works until the first kind stops running and the supply of people who know your name quietly dries up, which shows up months later as a decline nobody can trace.

The practical habit is to keep one clear key event that genuinely matters, read it against the platform reports rather than instead of them, and ask new customers how they found you. That last answer is unscientific, arrives in small numbers, and regularly contradicts the dashboard in ways the dashboard cannot see.