“Which attribution model should we use?”
So someone pulls up an attribution tool’s settings. First-click. Last-click. Linear. Time-decay. U-shaped. (GA4 itself now offers only data-driven and last-click; Google retired the rest in 2023.) Each one moves credit, and money, to a different channel.
And everyone in the room has a favorite.
An attribution model is a rule for splitting credit across the clicks you recorded. It can make a report consistent. It can’t tell you, on its own, which click caused the order.
That limit is what makes the output usable. So the real question is which rule fits which job, and where each one misleads.
What is an attribution model, exactly?
Write down four things before you pick one:
- The decision the report will feed.
- The event record you actually have.
- The order system you’ll check it against.
- What you’ll do if the result moves.
If none of those are clear, changing the model usually just changes the argument.
In this guide, every “touch” is a recorded click, as in click-based tools like GA4. Some tools also count ad views or impressions, but any model can only split credit among the touches it actually recorded.
Let’s use one made-up path for every rule below. A shopper clicks a Reddit ad, then a Facebook ad, then an email, then a Google search ad, and buys.
What does each rule do with that?
Last-touch attribution: where does it mislead?
Last-touch (also called last-click) gives 100% of the credit to the final recorded click. In our path, Google search gets the whole order.
For a lot of brands, this is the model they already run. It’s what Shopify, GA4, or the order system records from the UTM or click ID on the click that lands and buys.
But it only answers one narrow question: which recorded click came right before the order?
It doesn’t tell you which earlier channel built the demand. Search often sits at the end of a path sometimes because people search for a brand they already heard about somewhere else. Read last-touch as “search made this sale” and you’ll starve whatever put the brand in their head.
Where it fits: checking a tool’s view against your order record. If your order system says a sale was Facebook last-click, an MTA’s last-click view should say the same. Start there before you trust anything fancier. The ecommerce attribution-tool evaluation guide walks through that check.
First-touch attribution: where does it mislead?
First-touch is the opposite end. It’s like crediting the whole meal to the appetizer.
It gives 100% of the credit to the first recorded click. In our path, Reddit gets the order.
That’s useful for one job: seeing which source first brought a buyer into the record. It can help with prospecting and creative reviews.
It misleads in two ways.
First, it hands every later outcome to whatever was recorded first. The Facebook ad, the email, and the search all did nothing, as far as this rule is concerned.
Second, it depends on knowing the shopper is the same person over time. That’s hard. Cookies get cleared. Safari and other browsers shorten or block them. People switch devices. When the system mistakes a returning buyer for a new visitor, it records a fresh “first” touch that isn’t really first.
Without first-party event capture (site events you record on your own domain), too much traffic has “goldfish memory.” First-click can collapse into last-click, and you learn nothing new.
Linear attribution: where does it mislead?
I have one caution with linear: equal isn’t the same as fair.
Linear splits credit evenly across every recorded click. Four clicks in our path, so each gets 25%.
It’s easy to explain, and it makes a long path easier to review.
But the even split is a choice, not a finding. Nobody measured that the email mattered exactly as much as the Reddit ad. And a channel that racks up lots of cheap clicks in a path collects more slices just by showing up more often.
Label it as a convention, and never present its weights as measured effect.
Time-decay attribution: where does it mislead?
Does a click closer to the purchase deserve more credit?
Time-decay says yes. It gives more weight to recent clicks and less to older ones. In our path, Google search gets the biggest share, then email, then Facebook, and Reddit gets the least.
That can make sense for a short buying cycle. It’s a way to look at recent returning activity.
The catch is that it leans the same way last-touch does, just more gently. Channels that start demand sit early in the path, so they get discounted by design. And with a long consideration cycle, a high-AOV product (average order value) for example, the early click may have done more of the work than the model can show, yet it gets the least credit.
So what if you care about both ends?
Position-based attribution: where does it mislead?
My caution here: the split looks smart, but someone picked it.
Position-based (also called U-shaped) gives most of the credit to the first and last clicks and spreads the rest over the middle. A common setup, and the one Google Analytics used before it retired the model, is 40% first, 40% last, and 20% split across the middle. In our path, Reddit and Google search get 40% each, and Facebook and email get 10% each.
It respects both the click that started the path and the click that closed it. That’s appealing.
But it has both blind spots at once. It still needs reliable identity to know what was really first. And the middle gets squeezed, so a channel that nurtures people between discovery and purchase looks weaker than it may be.
Every rule-based weight is a policy you chose. Write it down like one.
What can’t any attribution rule fix?
Missing data.
No rule fills in event capture you never recorded. No rule matches identity across every device. No rule sees an ad that was watched but never clicked. TV, podcasts, and radio have little or no click trail, so click-based models will usually credit them with nothing.
And no rule proves incrementality. If the claim is “this channel created more total orders,” use an incrementality test when you can run one. If the question is how to split budget across many channels over time, an MMM (marketing mix model) may give the wider view.
If no single rule is right, how should you use them?
How should you use attribution models?
Run more than one. Then:
- Look across models. Each rule moves credit around to show a different idea about a messy customer journey. Use them to form theories.
- Look for signals that agree. Before giving a channel a test slot, check for positive results across first-click, last-click, U-shaped, and time-decay views. Also check brand-new visitors and a higher share of new-to-file customers among claimed orders.
- Test the theory. None of those clues proves incrementality. Returning customers can still be incremental, even after a lag. The clues tell you what to test first.
- Write a short note for every model you report. Its purpose, scope (channels and lookback window), data limits, the exact rule, and the claims it must not be used to make.
If that note feels hard to write, the model probably isn’t ready to steer spend. A custom MTA should show those gaps through reconciliation, not hide them behind a nicer chart.
No model owns the truth. Use several to build a theory, then test it.
FAQs
Which attribution model is most accurate?
None of them, on their own.
Picture this: last-touch says search drove the quarter, first-touch says Reddit did, and linear says it was a bit of everything. Each one is right about its own rule. The only way to check what actually caused sales is a test, ideally a geo test.
Is last-touch attribution still useful?
Yes. It’s the cleanest way to check that a tool’s numbers match your order record. Just don’t use it to decide which channel built demand.
What’s the difference between rules-based and data-driven attribution?
Rules-based models (first-touch, last-touch, linear, time-decay, position-based) split credit by a formula someone picked. Data-driven, or algorithmic, models set the weights from patterns in your data. Both still work from recorded touches (usually clicks), so both still describe what was observed, not what was caused.
Do I need multi-touch attribution?
It depends on your data and your decision. Start with how to choose an ecommerce attribution tool.