Attribution models: which model for which decision
The choice is determined by the decision you are making: last click tells you what closes the sale, first click shows what started it, and the linear model spreads the credit across the whole path. The attribution model is the rule that decides which interaction a sale is recorded against, when the person has seen several ads before buying. It determines which channel looks profitable and where the budget goes next month. The same 100 sales, viewed through different models, distribute the credit quite differently: last click often lifts remarketing and branded queries, and switches off the campaigns that find new customers. That is why the model is not a technical setting but a frame through which you read the numbers before moving budget between channels.
My name is Hristiyan Panov — a performance marketing consultant with over 5 years’ experience and over €2 000 000 managed advertising budget for over 90 companies on Google and Meta. Below is what each model does and which one to use for which decision.
What are the main models?
The models are a finite set of rules for distributing the same credit. The difference between them is clearest when you open the same period through two models in turn and compare the conversion columns.
- Last click — all the credit goes to the last interaction. The most widespread and the most misleading, because it is also the default setting.
- First click — all the credit goes to the first. It shows which channel discovers new people and lifts the campaigns at the top of the funnel.
- Linear — the credit is split equally between all the steps, no matter how far they are from the purchase.
- Time decay — those closer to the purchase get more, and interactions from weeks earlier weigh less and less.
- Data-driven — the platform distributes credit using its own model, built from your data.
Where the setting lives differs on each platform. In Google Ads the model is set on the conversion action, not on the campaign, so the change applies to every campaign that reports that action. In GA4 the choice is made at property level. In Meta, the attribution windows in the report itself play the role of the model.
Why is last click misleading?
Because it rewards the channel that picks the fruit. Branded search on Google is almost always the last step — the person has already decided and is simply searching for your name. The channel that created that decision stays invisible.
It is recognised by the same set of signs. The branded campaign shows the best ROAS and the lowest CPA in the account, but total revenue stays flat. The share of queries containing the brand name in the search terms report is growing. Campaigns for a cold audience show a high CPA and look like the first candidate to be switched off. The assisted conversions and path-to-purchase reports show those very campaigns at the start of the path.
The result is predictable: the budget moves towards the bottom of the funnel, new customers decline, and after a few months growth stalls while the reports look excellent. The check is simple — follow the share of new versus returning customers alongside ROAS, not ROAS alone.
When is first click more useful?
When the question is where the new people come from. First click is not more true than last click — it simply answers a different question. Use it as an additional view, not as a replacement.
The practice that works: look at the same number through two models and pay attention to the difference, not the value. The order is as follows. Fix one period and one conversion action. Pull the conversions by channel under last click. Pull the same conversions under first click. Rank the channels by the difference between the two values.
The reading is mechanical. A channel that wins under first click does the discovery work and is rarely judged fairly by the last step. A channel that loses under first click is a closing channel and is usually already well covered. Check the ranking against the share of new customers for the same period before moving budget.
What does the data-driven model do?
It distributes credit according to which combinations of interactions actually lead to a purchase in your account. It requires volume — with few conversions the model has nothing to learn from and the result is unstable.
Before you trust the distribution, check the foundation beneath it. The conversion action must be a single, clearly defined one, not several overlapping actions counting the same order. It must be in the primary conversions column, because only that one feeds machine learning. The tag must fire once per order, without duplication when the thank-you page is reloaded.
An important limitation: Google’s model sees only Google’s interactions, and Meta’s model only Meta’s. Neither sees the whole picture. Switching to this model also recalculates the history backwards, so comparisons with previous months change.
Why do the platforms report more than you have sold?
Because each one credits itself with its share of the same orders. Their sum regularly exceeds the real revenue — not because they lie, but because each one measures in its own world with different attribution windows.
The discrepancy is measured rather than debated. Take one closed month. Pull the revenue from the cash books or from the store’s admin panel. Reconcile on the same date — order date, not payment date, because the two give different totals. Put the revenue reported by Google, by Meta and by GA4 next to it. The ratio between the platform figure and the real revenue is the coefficient with which you read the reports from here on.
Account for returns as well: the platforms record the order at checkout, while the till corrects it later, when the goods are returned or the order is refused on delivery.
How do you make a decision despite the differences?
The rule is simple: every decision is made with the number that is entitled to answer its question.
- Within a channel — use the platform’s model and compare campaign with campaign, over the same period and with the same conversion action.
- Between channels — use the overall calculation: total revenue against total spend. Platform ROAS figures are not added together.
- For growth — look at the growth in new customers, not at the credit for the last order.
- If in doubt — pause a channel for a short time and see what happens to total revenue. This is the most honest test.
The pause test has conditions without which it is worthless. One channel is paused, not two at once. The period is long enough to cover the usual path to purchase for you. The total revenue is compared, not the revenue of the paused channel. The dates of pausing and of restarting are recorded, because promotions and seasonality explain part of the movement.
Does consent to cookies make a difference?
Yes, and significantly. Part of the user’s path disappears on refusal, which distorts every model. Consent Mode and Enhanced Conversions recover part of what is missing, but not all of it.
The check goes through GTM. See whether the tags really wait for consent instead of firing when the page loads. See whether a state update is sent on confirmation and whether the tags fire after it. See whether the conversion tag receives the hashed data for Enhanced Conversions on the page after checkout.
The practical consequence is in reporting. A drop in conversions straight after a change to the banner is almost always a measurement drop, not a commercial one. One check tells them apart: if the revenue in the cash books is unchanged while the platform figures have fallen, measurement has been lost.
FAQ — Frequently asked questions about attribution
Which model is the most correct?
There isn’t one. Each answers a different question — the mistake is to expect a single model to answer them all.
Will my historical data change when I switch the model?
In most platforms, yes — they recalculate backwards. That is why a switch needs an explanation before anyone sees a “crash” in the report.
Which attribution window should I choose?
As long as the real path to purchase is for you. An impulse purchase and an expensive B2B service are not measured by the same yardstick.
Is a separate attribution tool worth it?
With several channels and a large budget — sometimes. With a small account the difference rarely justifies the cost and the complexity.
How do I compare Google with Meta fairly?
Not through their reports. Through the overall calculation and through a test that pauses one of the channels.
Conclusion
Attribution is not truth but a point of view. A useful decision comes not from choosing the “right” model, but from knowing what question you are asking and which number is entitled to answer it.
If your channels are fighting over the same credit, start with an independent review.
Request a quote for a digital audit →
Related articles: GA4 and Google Ads · MER · UTM parameters
