measurement · 8 min read
Attribution models: the same journeys, divided six different ways
An attribution model does not discover which channel caused a sale. It applies a rule for dividing credit among the touchpoints that happened to be recorded. This explains each model, shows what they do to one identical set of journeys, and states the limits none of them can escape.
Written by Zubair Afzal, FounderUpdated
An attribution model does not find out which channel caused a sale. It applies a rule for dividing credit among the touchpoints somebody managed to record.
Everything that follows comes from that sentence. If the model is a division rule rather than a discovery procedure, then no model can be more true than another in the way people usually mean, and choosing between them is closer to choosing an accounting convention than to establishing a fact.
That is not an argument for giving up. Marketing attribution models are genuinely useful: they make channel comparison possible, they force consistency, and they surface patterns nobody would see by eye. It is an argument for knowing what you are holding, so a report can be used for the questions it can answer and set aside for the ones it cannot.
Inputs
What a model receives before it divides anything
Every model is fed the same thing: a set of recorded journeys. Each journey is a list of touchpoints with timestamps, ending in a recorded outcome. Those touchpoints exist because something identified them — a tagged campaign link, a referring site, a recognised search source — and were grouped into channels by a rule somebody either configured or accepted as a default.
Three constraints are applied before the model ever runs. The lookback window decides how far back a touchpoint can be and still count, which means anything earlier is simply absent. Consent decisions determine whether a journey was recorded at all. And reporting thresholds can withhold detail on small segments, so the smaller the slice you examine, the more likely you are looking at an incomplete picture without being told.
The result is that a model is not dividing the journey. It is dividing the recorded fragment of the journey that survived those three filters. On a business with a short online path and clean tagging, that fragment is most of the story. On a business where the decisive conversation happened at a trade show or on a phone call, it is a minority of it, and the model will still divide it with perfect confidence.
The rules
Six models, and what each one is actually saying
Each rule embeds a theory about how buying works. Reading the theory is more useful than reading the name, because the theory is what you are asserting when you pick one.
| Dimension | The rule | The assumption it makes | Where it is defensible |
|---|---|---|---|
| Last click | All credit to the final recorded touchpoint before the outcome. | That the last thing before the purchase is the thing that caused it. | Short, single-session buying, and as a consistent baseline everybody understands. |
| First click | All credit to the first recorded touchpoint in the window. | That discovery is what matters and everything after it was going to happen anyway. | Evaluating activity meant to create awareness, read alongside another model rather than alone. |
| Linear | Credit split evenly across every recorded touchpoint. | That every touch mattered equally, which is almost certainly false and admits it. | A neutral starting point when nobody can justify a shape, and for seeing which channels are involved at all. |
| Time decay | More credit to touchpoints nearer the outcome, on a decreasing curve set by a half-life. | That influence fades with time, and recent contact matters more. | Considered purchases with a defined cycle, where the half-life can be set from real deal length. |
| Position-based | A large fixed share to the first and last touch, the remainder split between the middle. | That discovery and closing matter most, and the middle is supporting work. | Longer journeys where both ends are genuinely distinct activities run by different teams. |
| Data-driven | Credit estimated by comparing converting paths with non-converting ones. | That contribution can be inferred from the difference between the two populations. | Accounts with enough volume for the comparison to be stable, and audiences prepared to accept a rule they cannot check by hand. |
Worked example
One identical set of journeys, divided by each rule
A hypothetical illustration, not a measurement of anything. It exists so the mechanics are visible: the same journeys, the same outcome, six different pictures of who did the work.
| Dimension | Organic, day 21 | Paid social, day 14 | Email, day 7 | Brand search, day 0 |
|---|---|---|---|---|
| Last click | 0 conversions | 0 conversions | 0 conversions | 100 conversions |
| First click | 100 conversions | 0 conversions | 0 conversions | 0 conversions |
| Linear | 25 conversions | 25 conversions | 25 conversions | 25 conversions |
| Time decay | About 6.7 | About 13.3 | About 26.7 | About 53.3 |
| Position-based | 40 conversions | 10 conversions | 10 conversions | 40 conversions |
| Data-driven | Not derivable from one path | Not derivable from one path | Not derivable from one path | Not derivable from one path |
Limits
Four things no model fixes
These are properties of the whole approach rather than defects of any particular rule, which is why switching models never resolves them.
- The touchpoints nobody recorded.
- A recommendation from a colleague, a podcast mention, a conversation at a conference, a link shared in a private message. These frequently decide the purchase and appear in no journey. Every model divides the recorded fragment, and the more of the real journey happens offline, the more precisely the model describes a minority of what occurred.
- Consent and thresholds thin the data unevenly.
- Journeys are only recorded where a visitor allowed it, and reporting systems can withhold detail on small segments to protect identity. Neither effect is distributed evenly across channels or audiences, so the shape of what survives is itself biased, and no division rule applied afterwards can restore what was never collected.
- The systems do not agree on what a conversion is.
- Different platforms use different lookback windows, count either every conversion or one per click, and some report modelled outcomes where a direct link could not be observed. Attribution runs on top of that inconsistency. Reconciling models is impossible while the underlying counting rules differ, and knowing the rules is the only way to interpret the disagreement.
- No model addresses the counterfactual.
- Every model answers who was present. None of them answers what would have happened if a channel had not run, and that is the question every budget decision actually asks. Only a withholding test gets at it, and no amount of modelling substitutes for one.
Practice
How to work with models without being misled by them
None of this requires new software. All of it requires somebody to write things down and stick to them.
- Pick one model as the reporting standard and state it on every report, so nobody has to ask which set of rules produced the figures.
- Treat a model change as a scheduled event: restate the history, re-baseline the targets, and tell everyone who has a number in a plan.
- Read a second model alongside the standard one, not to average them, but because the gap between them tells you which channels are early and which are late.
- Write the lookback window on the report. A thirty-day window and a ninety-day window produce different histories from identical activity.
- Reconcile total attributed conversions against what the business actually recorded, and publish the variance rather than hiding it.
- Keep an unmodelled view: raw channel sessions and raw enquiries, unallocated. It is the check that tells you when a model change is doing something odd.
- Ask customers how they heard about you, and treat those answers as a separate, imperfect witness rather than as data to reconcile against.
Questions
What people ask once the reports disagree
Which attribution model should we use?
The one you will keep, that your leadership team can understand, and that you re-baseline your targets against when you adopt it. Consistency is worth more than sophistication for most businesses, because the value of the report is comparison over time.
If the buying journey is short and mostly single-session, last click is defensible and simple. If it is long and multi-touch, a position-based or data-driven approach describes it better. Neither choice makes the underlying data more complete.
Is last click really that bad?
It is not wrong, it is narrow. It answers one legitimate question precisely: which channel was the final recorded touch. The problem is that people read it as an answer to a different question, which is which channel caused the sale.
The predictable consequence is that channels appearing late in the journey look efficient and channels creating demand look wasteful. If last click is the reporting standard, that bias should be written into the commentary, not left for somebody to discover during a budget cut.
What is data-driven attribution actually doing?
Comparing paths that converted with paths that did not, and estimating each touchpoint’s contribution from the difference. It is not applying a fixed proportion, which is why you cannot reproduce its output by hand from a single journey.
That makes it more statistically defensible than a fixed rule and considerably harder to explain. It also needs enough data to work with, and its results shift as the underlying data shifts, which can look like instability to anyone expecting a constant.
Why do our numbers change when we switch models?
Because the model is the rule that produced the numbers. Change the rule and the history changes with it, often substantially for upper-funnel channels.
The operational mistake is switching without re-baselining. Targets, budget splits and sometimes commission plans were all set under the old rule, so the switch has to be a scheduled event with a restated history rather than a settings change somebody makes on a Tuesday.
Where do offline and word-of-mouth influence appear?
Nowhere. No model can allocate credit to a touchpoint that was never recorded, and the conversation that actually decided the purchase usually happened somewhere no tag can reach.
This is the limit worth stating out loud in any report. Models redistribute what was seen; they do not discover what was missed. Businesses with heavy referral or offline influence should treat the whole attribution output as a partial view and lean harder on withholding tests and on asking customers directly.
Work out which model your reporting should stand on
The right answer depends on how your customers actually buy and what your leadership team needs to be able to explain. It is a short conversation, and it is worth having before the next model change rewrites a year of history.
Related
Where to go next
Last updated