Six attribution models, one set of orders: how to read the disagreement.

A customer clicks a Meta ad on Monday, searches for the product on Google on Friday, and buys through an email the following Tuesday. Which channel earned the sale?

Every answer you have ever been given to that question came from a rule. Last click says email. First click says Meta. Linear splits it three ways. None of them is a measurement of what caused the sale. Each is a convention for dividing credit among the touches somebody happened to observe.

That sounds like a reason to distrust attribution. It is actually a reason to use more of it. One model gives you one opinion, stated with false confidence. Several models, run on the same orders and read side by side, give you something much more useful: a map of where the credit is solid and where it depends entirely on the rule.

What an attribution model is, and is not

An attribution model takes a set of observed touches before a conversion and assigns each one a share of the credit. The rule might be positional, such as all to the last touch, or temporal, such as more to recent touches, or learned from data, such as comparing the paths of people who bought with the paths of people who did not.

What every model shares is that it only redistributes credit among touches that happened. It cannot say whether the sale would have happened anyway. A customer who was always going to buy, and clicked a brand search ad on the way, still hands that ad some credit under every model. That question belongs to incrementality, and it needs a different tool.

One journey, six models

Take the journey above, a $300 order, with the Meta click eight days before purchase, the Google search click four days before, and the email click on the day. As arithmetic:

The rule-based models are simple enough to compute by hand. Time decay gives each touch a weight that halves for every seven days between it and the purchase, then scales the weights to the order value. Position based gives 40% to the first touch, 40% to the last, and shares the remaining 20% among everything in between.

Data-driven attribution cannot be computed from one path. It compares converting and non-converting journeys across the account and estimates how much each kind of touch changes the chance of a sale. It is usually the most sensible single model when there is enough data. It is also the hardest to audit, which is one reason not to rely on it alone.

Why you may only see two of them

In 2023 Google removed first click, linear, time decay, and position-based attribution from Google Ads and Google Analytics 4. Both now offer data-driven and last click. The removal made the reports simpler and the comparison harder, because the fastest way to see a channel’s role is to watch how its credit moves between rules.

There is a second limit. Each ad platform’s own model sees only its own touches. Meta cannot see the Google search, and Google cannot see the Meta view. So comparing Meta’s report with Google’s is not comparing two models. It is comparing two partial views of the path. To compare models properly, you need the whole path in one place, which in practice means first-party tracking on your own domain.

A month, five rules

Here is an illustrative month of $200,000 in tracked revenue, credited under the five rule-based models. The numbers are arithmetic, chosen to show the patterns, not any client’s result. Every column adds up to the same $200,000. Only the split changes.

Read the Range column first. Google Ads sits between $60,000 and $70,000 whatever rule you pick. That credit is earned in the sense that matters for budgeting: no reasonable convention takes it away. Organic and direct barely move either.

Meta and email are different. Meta’s credit more than doubles from last click to first click. Email’s falls by two-thirds in the same move. That is not noise. It is a description of their jobs. Meta opens journeys in this account, and email closes them. Under last click, Meta looks like the weakest paid channel. Under first click, it looks like the strongest. Neither view is the truth. Together they tell you what the channel does.

How to read the disagreement

1. Find the channels with a narrow range. Their credit does not depend on the rule, so budget decisions about them can lean on attribution with reasonable confidence.
2. Find the channels with a wide range, and look at the direction. Credit that rises toward first click marks an opener. Credit that rises toward last click marks a closer.
3. Never cut an opener on last-click evidence alone. This is the most expensive misreading in paid media. Cutting the channel that starts journeys starves the closers of people to close, and the damage shows up weeks later in channels that look unrelated.
4. Be skeptical of closers that look too good. Brand search, retargeting, and email capture demand that something else created. High last-click credit for them is partly a measure of the brand.
5. Settle the big, wide-range questions with a test. When a large budget depends on a channel whose credit swings with the rule, a geo holdout or a platform lift study answers what no model can.

The lookback window moves the answer too

A model is a rule plus a window. The window decides which touches are old enough to be forgotten. Shorten it, and the touches that happened early in a long journey fall out of the path, so credit shifts toward whatever came last. Lengthen it, and openers regain some of what they lost.

Take the journey above with a seven-day lookback. The Meta click, eight days before the purchase, is no longer in the path. Every model now divides the $300 between Google and email, and first click hands the whole order to Google. Nothing about the customer changed. Only the window did.

• →Use one window across every model you compare, or the columns are measuring different things.
• →Choose it from the business’s real purchase cycle, not from a platform default. A considered purchase with a two-week research phase needs a longer window than an impulse buy.
• →When you change the window, expect openers to move most. Treat that move as a consequence of the setting, not a change in performance.

Checking data-driven attribution against the rules

Data-driven attribution is a good default and a poor sole authority. It needs enough conversions and enough variety in paths to learn from, and when it has them it usually produces sensible splits. But it is hard to inspect, and the platforms that run it are also the platforms being judged by it.

The rule-based models give you a way to check it. Lay its credit for each channel beside the range from the rule-based models. When data-driven credit sits inside that range, it is choosing a point among reasonable conventions, and that is easy to accept. When it sits well outside the range for a channel, for example crediting Meta more than first click does, it is making a claim no simple rule supports. That does not make it wrong. It does make it the thing to investigate before moving money on its evidence.

What the models cannot see

Every model runs on observed touches, and some touches are never observed. Ad views without a click rarely reach your own tracking. A customer who researched on a phone and bought on a laptop looks like two people unless they identified themselves on both. Visitors who decline tracking consent leave gaps in the path. And no model can see a friend’s recommendation, a podcast, or a shop window.

That is why the models should be read against the store, not instead of it. The total they are dividing is tracked revenue, which is smaller than the store’s net revenue. If the gap between the two grows, the models are dividing a shrinking share of reality, and the right response is to fix tracking before trusting any split. The full reasoning is in why Meta, Google, GA4, and Shopify report different revenue.

Where to run them

Running several models properly needs three things. First, one path per customer that includes every channel, which the platforms cannot provide because each sees only itself. Second, the same definitions for every model, the same conversion, the same lookback window, and the same revenue, so that the only thing changing between columns is the rule. Third, the order value from the store, so the credit is split on money the business actually took.

That usually means first-party tracking on your own domain, with purchases captured server side so they survive browsers and ad blockers. It is how we run it for clients: the platform behind our agency computes all six models on each client’s own tracking data and shows them next to each other, so the conversation starts from where they agree. What to collect for that, and where to send it, is covered in first-party data for paid media.

A monthly routine

1. Check that tracked revenue is a stable share of store net revenue. If it moved, investigate tracking before reading anything else.
2. Lay out credit by channel under each model, with the range beside it, exactly as in the table above.
3. Compare each channel’s range with last month. A channel whose role changed, for example from opener to closer, usually reflects a change in creative, audience, or budget worth understanding.
4. Make budget moves on narrow-range channels with confidence, and on wide-range channels only in small steps or behind a test.
5. Write down the decision and the models that informed it, so next month’s review can check the reasoning, not just the result.

The routine also changes the conversation with whoever runs the budget. Instead of arguing about which dashboard is right, the discussion becomes which channels have earned their credit under every rule, which ones play a role the rules disagree about, and which of those disagreements is large enough to be worth a test. That is a conversation a finance lead can follow, and one where a bad decision is much harder to defend.

None of this makes attribution certain. It makes it honest. And an honest map of where credit is solid is worth far more than a confident number that happens to flatter whoever chose the model.