A buyer finds a jacket through a social ad, reads a sizing guide, leaves, returns through an email, and eventually buys after a branded search. Nothing about that purchase changes when you switch the attribution model. Only the story in your report changes.
In a hypothetical journey, those interactions might unfold over a week:
Monday: A social ad introduces the jacket. The buyer visits the store but does not buy.
Wednesday: They return to read the sizing guide and subscribe for updates.
Friday: An email reminds them about the jacket. They look again.
Sunday: They search for the brand, click a paid search result, and complete the order.
Now bring that one order into a budget meeting.
First-click attribution gives the credit to Monday's social ad. Last-click attribution gives it to Sunday's paid search. The email and sizing guide receive nothing under either single-touch model, even though both appear in the observed path.
Two reports. One purchase. Four interactions. The disagreement is not necessarily an error; it is what those models were built to produce.
Follow the order, not the winner
First-click asks which recorded interaction began the journey. Last-click asks which recorded interaction came immediately before conversion, subject to the model's rules. Neither question is the same as “Which interaction caused this person to buy?”
That difference matters. Under first-click, social looks like the winning channel in our example. Yet the model cannot establish that the buyer would have purchased without the email or search result. Under last-click, paid search looks decisive. Yet the report cannot tell us whether the buyer already intended to buy before searching.
The credit is an accounting choice, not a controlled experiment.
It is tempting to describe first-click as measuring discovery and last-click as measuring persuasion. Even that distinction needs care. The first interaction you can observe may not be the person's first encounter with the brand. An untracked recommendation, earlier visit on another device, or offline exposure could come before it.
Likewise, the last interaction may capture existing intent rather than create it. A customer typing a brand name into search may already have made the choice. Recording the final click does not reveal how much additional demand that click created.
Before comparing the two reports, make sure they are describing the same population. Use the same conversion definition, dates, attribution window, and inclusion rules. Check how direct visits are handled. A system that excludes direct traffic can credit an earlier eligible interaction rather than the literal final visit. If settings differ, some of the apparent argument between models is actually an argument between configurations.
This is also why recorded revenue and attributed revenue can be different reporting concepts. Model credit may be assigned or redistributed without changing the underlying order. Establish which number you are comparing before treating a difference as missing sales.
Our hypothetical buyer offers another useful observation: the middle interactions are visible even when single-touch credit ignores them. An email may repeatedly appear between discovery and purchase without winning either report. That is a reason to investigate its role, not proof that it deserves a particular share of the budget.
What changes in the budget meeting?
Suppose someone proposes moving spend from social to branded search because the last-click report looks stronger. Ask what would happen to the customer path if fewer new people encountered the product in the first place.
Do not assume the answer. Instead, separate what the report describes from what the team is predicting. The numbers describe where eligible conversions were credited. The prediction is that moving the budget will produce a better business outcome. Those are different claims and need different evidence.
Looking at first-click alongside last-click can reveal that a channel has a different role at different stages. It can also reveal a tracking gap, a short observation window, or a customer group whose behavior deserves closer inspection. Segmenting by new versus returning customers may help explain the difference, provided that definition is consistent and the underlying data supports it.
For the jacket order, a useful next question might be whether new customers commonly start with social and finish with search. Another might be whether repeat buyers take a shorter path. Neither question requires declaring one attribution model correct for every decision.
If the evidence warrants changing a budget, make the prediction explicit before acting. For example: reducing a discovery channel could lower the number of new customers entering observed journeys, even if the closing channel's attributed performance remains attractive for a while. Decide which outcome would support or weaken that prediction, and over what period it is reasonable to evaluate it.
A well-designed experiment can provide stronger evidence about incremental impact than model credit alone, when experimentation is feasible. Multi-touch attribution can distribute credit across more observed interactions, but adding touchpoints does not automatically turn the result into causal evidence. More detailed allocation remains allocation.
Return to that Sunday order. Paid search did not become a better channel when you changed a dropdown, and social did not become a worse one. The buyer still followed the same path.
The most productive use of the disagreement is to ask a sharper question about that path—and choose a way to investigate it—before changing the next budget.