Attribution models are an argument about credit, not a source of truth
First touch, last touch, linear, time decay, U-shaped, data-driven. Six ways to divide a number that nobody has verified. Here is what each model is actually claiming, when the choice genuinely matters, and why arguing about it before you can match a sale is arguing about the wrong thing.
Attribution model debates have a particular quality: enormous energy, high confidence, and almost no evidence on either side. Somebody advocates a U-shaped model, somebody else prefers time decay, a vendor mentions data-driven, and the meeting ends with a decision that will shift millions of pounds of reported credit between channels. Nobody in the room can say what makes one of them right.
An attribution model is not a measurement instrument. It is a rule for dividing credit among things you have already decided were involved.
That is not a reason to ignore the choice. It is a reason to make it deliberately, understand what you are choosing, and stop treating the output as a fact about the world.
What a model is actually doing
An attribution model takes a sale, a list of recorded interactions preceding it, and applies a fixed or learned rule to split the revenue among them. The rule is a policy decision. It cannot be validated, because the counterfactual — what would have happened without a given touch — is never observed.
This is the part that gets lost. You can validate a match: this sale really did come from this enquiry, here are both records. You cannot validate a weighting. There is no experiment in which the customer buys and simultaneously does not buy, so there is no ground truth for the claim that the display impression deserved eighteen percent.
The one family of methods that does address causality — geographic holdouts, incrementality tests, randomised experiments — works by deliberately withholding marketing from some people and comparing outcomes. That is a different activity from attribution, it costs real money in foregone revenue, and it answers a narrower question very well.
The six models, plainly
The standard models are first touch, last touch, linear, time decay, position-based and data-driven. Each encodes a different belief about what causes a purchase, and each one flatters a different part of your marketing.
| Model | The claim | Flatters | Best suited to |
|---|---|---|---|
| First touch | Discovery causes the sale; everything after is delivery | Awareness, content, organic search | Businesses whose problem is being found at all |
| Last touch | The final push causes the sale | Brand search, retargeting, direct | Short cycles, impulse or urgent purchases |
| Linear | Every touch contributed equally | Channels appearing often in long paths | Complex journeys where no touch obviously dominates |
| Time decay | Recent touches matter more | Late-funnel and closing activity | Cycles where urgency builds toward a decision |
| Position-based | The first and last matter most; the middle assists | A balance of discovery and closing | Considered purchases with a clear start and end |
| Data-driven | Credit follows observed statistical contribution | Whatever the data supports, in principle | High volume, well-instrumented, mostly digital paths |
Read that table as a list of opinions rather than a list of methods, because that is what it is. Linear attribution does not discover that every touch mattered equally; it assumes it. Time decay does not detect recency effects; it imposes them.
The argument nobody in the meeting is having
Before a model can divide credit among touches, the touches have to exist in your data. In most mid-market businesses a large share of revenue arrives through channels that record no path at all, so the model is operating on a biased subset of journeys.
This is the flaw that makes most model debates premature. If sixty percent of your closed revenue came in by phone, referral or repeat purchase, then your multi-touch model is carefully apportioning credit across the forty percent that happened to be clickable. The result is not a view of your marketing; it is a very detailed view of your website.
Worse, the bias is not random. The channels that generate observable paths are the ones you have instrumented, which are the ones you already believe in, which means the model tends to confirm the investment that produced the instrumentation. That is a hard loop to see from inside.
Choosing a sophisticated model while half your revenue is unmatched is like arguing about the decimal places on a number whose first digit you have not established.
When the choice genuinely matters
The model choice matters most when journeys are long, multi-channel, and well observed — and matters least when most sales have one recorded touch or none. In a business where the median path has a single interaction, every model produces the same answer.
That is a genuinely useful test and it takes ten minutes. Count how many recorded touches your matched sales actually have. If the median is one, the model debate is theoretical for your business and you can pick last touch, write it down, and move on to something that will change a decision.
| Median recorded touches per sale | Effect of model choice | What to do |
|---|---|---|
| 1 | None — all models agree | Pick one, document it, stop discussing it |
| 2 to 3 | Modest, mostly between two channels | Position-based or last touch; revisit annually |
| 4 or more, mostly digital | Large, and worth reasoning about | Consider data-driven, but insist on auditability |
| 4 or more, mixed offline | Large and mostly unmeasurable | Fix observability before modelling |
Data-driven attribution and the audit problem
Data-driven models are more principled than fixed rules and considerably harder to defend, because the weights are produced by a model you usually cannot inspect, using data you do not hold, in a system operated by a party with an interest in the outcome.
That last clause is the uncomfortable one. When an advertising platform's own data-driven model assigns credit to that platform's own inventory, the arithmetic may be perfectly sound and the arrangement is still one no auditor would accept in any other context.
This does not make the output useless. It makes it a vendor-reported metric, which is a legitimate category of information with well-understood handling: useful for optimising within the platform, not suitable as the basis for reporting revenue to a board.
Pick a model the way you pick an accounting policy
Choose a model for its properties — stability, explicability, resistance to gaming — rather than for its accuracy, which cannot be assessed. Then hold it constant long enough to trend, and change it only in public with the old series restated.
Accounting has been dealing with this problem for a century and has arrived at a workable answer: you cannot make the choice objectively true, so you make it explicit, consistent, and disclosed. Depreciation schedules are not discoveries about the world. They are policies applied uniformly so that periods can be compared.
- Choose the simplest model your journey data can support.
- Write down the choice, the date, and the reason, in one paragraph.
- Hold it for at least four reporting periods.
- If you change it, restate the prior periods on the new basis and show both series once.
- Never change it in the same meeting where the results are first seen.
The model matters less than the bucket you refuse to fill
Whatever model you choose, the most consequential number on the report is the revenue you could not attribute at all. A model divides what you matched; it says nothing about what you missed, and a tool that hides the gap makes every model look better than it is.
You can run linear attribution on sixty percent coverage and get a defensible report, provided the other forty percent is on the page. You cannot run the most sophisticated model in the world on the same data, hide the gap, and get anything honest, because the sophistication is being applied to a sample nobody has characterised.
CloseRev deliberately does not ship a menu of attribution models. It matches sales to sources, shows what matched and what did not, and lets you see the specific records behind both. Weighting a journey you cannot observe is the step where the honesty usually leaves.
The models nobody names, which run most businesses
Two informal models do more real work than the six textbook ones: the salesperson's opinion, recorded in a source field at intake, and the last thing the customer says when asked how they heard about you. Both are widely used, rarely acknowledged, and worth understanding on their own terms.
The self-reported source is genuinely valuable and genuinely unreliable, in a specific and predictable way. People misremember, they name the most recent or most memorable touch, and they compress a six-month journey into one word. But it captures things no tracking can: the conversation at a conference, the recommendation from a friend, the van they saw on a driveway.
The right treatment is to keep it as its own field, compare it against the matched source, and treat disagreement as information rather than error. When a customer says word of mouth and your data says paid search, both are probably true, and the interesting question is which one you would have lost the sale without.
What you should not do is use self-reported source as the primary attribution and present it as measurement. It is survey data collected by an untrained interviewer at an awkward moment, and it should carry the confidence that description implies.
Incrementality, and why it is a different question
Incrementality testing asks whether marketing caused revenue that would not otherwise have happened, by withholding it from a comparable group. It is the only method here that addresses causation, and it answers a much narrower question than attribution does.
The trade is real. A geographic holdout can tell you with reasonable confidence that a channel produced a measurable lift, but it takes weeks, costs foregone revenue in the held-out region, and gives you one answer about one channel over one period. Attribution gives you a complete, immediate, and much weaker picture of everything at once.
Sensible teams run both and use them for different purposes: attribution as the operational picture that gets reviewed monthly, incrementality as the periodic audit that keeps the operational picture honest. Treating either as a replacement for the other is where the money gets wasted.
A worked example of how much the choice moves
On the same set of journeys, switching between models routinely moves a channel's apparent contribution by a factor of two or more. The revenue did not change; only the rule for dividing it did.
| Channel | First touch | Last touch | Linear | Time decay |
|---|---|---|---|---|
| Organic search | 1,420,000 | 610,000 | 980,000 | 790,000 |
| Paid search | 690,000 | 1,340,000 | 1,010,000 | 1,180,000 |
| Paid social | 880,000 | 340,000 | 620,000 | 500,000 |
| 210,000 | 910,000 | 590,000 | 730,000 | |
| Total matched | 3,200,000 | 3,200,000 | 3,200,000 | 3,200,000 |
Look at email. Under first touch it is a rounding error; under last touch it is the second largest channel. A team that switched models mid-year without restating would see email appear to quadruple and would very reasonably conclude that something had worked. Nothing had.
This is why the discipline of holding the model constant matters more than the choice itself. The absolute numbers are policy artefacts. The changes over time, on a constant policy, are the signal.
What to say when somebody asks which model you use
Say which model, why you chose it, what share of revenue it is applied to, and what happened to the rest. Four facts. A model named without those is a label rather than an answer.
In practice this reframes the conversation productively. The question is usually asked as a competence check — does this person know the vocabulary — and answering with coverage rather than jargon moves it to the thing that actually determines whether the number is worth anything.
It also, quietly, tends to end the debate. Once everyone can see that thirty-eight percent of revenue has no recorded source, nobody wants to spend another hour on whether the middle touches should get twenty percent or thirty.
There is one more reason to answer with coverage rather than with a model name. Model names are portable between businesses; coverage is not. Two companies both running position-based attribution can be doing work of completely different quality, and the only way to tell them apart is to ask what share of revenue each one is actually able to trace. That is the number that separates a measurement practice from a vocabulary.
If you take one operational habit from this article, make it that: whenever an attribution figure is presented anywhere in your business, the coverage figure is presented beside it, in the same font, without being asked for.
Nobody has ever improved a business by switching from time decay to U-shaped. Plenty have improved one by finding the intake system that was never being exported.
Questions people actually ask
- Which attribution model is the most accurate?
- None of them, in the sense usually meant. Every model is a rule for dividing credit among touches, and no rule can be validated against a known answer because the true cause of a purchase is unobservable. The honest framing is that a model is a policy, chosen for what it encourages, not a measurement that can be right or wrong.
- What is the difference between first-touch and last-touch attribution?
- First touch gives all credit to the earliest recorded interaction, so it rewards discovery and demand creation. Last touch gives all credit to the final interaction before purchase, so it rewards closing and capture. Both are single-touch models and both are wrong in the same way: they assign one hundred percent of the cause to one event in a sequence.
- Is last-click attribution still used?
- Widely, yes, despite two decades of criticism, because it is simple, cheap, stable, and everybody understands it. It systematically undercredits upper-funnel activity, and knowing that, some teams reasonably keep it as a consistent operational yardstick rather than a truth claim.
- What is data-driven attribution?
- A model that assigns credit statistically by comparing paths that converted with paths that did not, rather than by a fixed rule. It is more defensible in principle and much harder to audit in practice, because the weights come out of a model whose inputs you generally cannot inspect.
- Do I need a multi-touch model?
- Only if you can reliably observe multiple touches per customer. If most of your revenue arrives by phone, referral or offline, you cannot see the path, so a multi-touch model is dividing credit among the small share of journeys that happen to be visible. That is worse than a single-touch model applied honestly.
- Should I change attribution models if the numbers look bad?
- No, and the temptation to is the strongest argument for writing the choice down in advance. A model changed after seeing the results is no longer a measurement policy, it is a way of producing a preferred answer, and every trend line either side of the change becomes meaningless.