Path attribution models: last click, first, linear, position-based and time-decay
Splits the value of orders between channels in five ways from a table of customer paths: last click, first click, linear, position-based 40/20/40 and time-decay by the age of a touch. Shows each channel’s value in each model and the gap between them: where it is large, the assessment of the channel depends on the choice of model and cannot be decided by it alone. Attribution divides credit but does not prove causality; the proof of causality is a test against a control.
At a glance
- What question does it answer?
- What value each channel gets in five models and how much the assessment depends on the choice of model.
- What data is needed?
- Customer paths: the sequence of channels before the order and the order value.
- How to read the result?
- Look for channels whose assessment depends strongly on the model: do not decide on them without an experiment. A channel that wins in all models is reliable. Example on the default values: Value of the paths: RUB 240; Paths: 3; Average touches per path: 2; Share of value in single-touch paths: 25%. The figures are illustrative: enter your own to see your result.
Why and how to reason
An order often passes through several channels, and different rules divide credit differently: last click, first, equally, 40/20/40, time-decay. The spread between models shows how much the assessment of a channel depends on the rule. In the example by the last click Email gets 180, Search 60; by the linear Search 93.3, Social 73.3, Email 73.3; by the position-based 100, 60 and 80.
When to use it
Apply when reviewing the marketing mix and before reallocating budget; causality needs a test against a control.
How to read the result
Look for channels whose assessment jumps between models: a decision on them without an experiment is unreliable.
Common mistakes
They take attribution for causality and compare channels over different observation windows.
What data is needed?
- Customer paths: the sequence of channels before the order and the order value.
- The half-life for the time-decay model and days between touches if there are no exact dates.
- The same channel name in all paths.
- An example: three paths (Search > Social > Email 100, Social > Email 80, Search 60): by the last click Email 180, Search 60; by the linear one Search 93.3, Social 73.3, Email 73.3 (illustrative numbers).
Inputs
| Field | Unit | Value in the example | Meaning |
|---|---|---|---|
| Half-life of a touch | days | 7 | After how many days the weight of a touch halves. For the time-decay model. The example value is illustrative. |
| Days between touches (if not set in the row) | days | 3 | An assumption about the distance between neighboring touches of a path when exact dates are unknown: the age of a touch = the number of touches after it × this value. The example value is illustrative. |
How it is calculated
| Step | Formula |
|---|---|
| Last and first click | all the value of the path to the last or the first touch |
| Linear | value ÷ number of touches to each touch |
| Position 40/20/40 | 40% to the first and last, 20% is split among the middle ones; two touches split 50/50, a single one gets all |
| Time-decay | weight = 0.5^(age of the touch ÷ half-life), weights are normalised to sum 1 |
| Total by channel | the sum of the shares over all paths |
Worked numbers
Default values from the app; in the interactive version you enter your own.
- Value of the paths
- 240 ₽
- Paths
- 3
- Average touches per path
- 2
- Share of value in single-touch paths
- 25 %There all models give the same.
- Channel with the largest gap between models
- Email · 180 ₽
- Largest gap, % of the value of the paths
- 75 %
- Last click
- First click
- Linear
- Position 40/20/40
- Time-decay
Data as a table
| Measure | Last click | First click | Linear | Position 40/20/40 | Time-decay |
|---|---|---|---|---|---|
| Search | 60 | 160 | 93.3 | 100 | 84.1 |
| Social | 0 | 80 | 73.3 | 60 | 66.5 |
| 180 | 0 | 73.3 | 80 | 89.5 |
Value by channel and model
| Channel | Last click | First click | Linear | Position 40/20/40 | Time-decay |
|---|---|---|---|---|---|
| Search | 60 | 160 | 93.3 | 100 | 84.1 |
| Social | 0 | 80 | 73.3 | 60 | 66.5 |
| 180 | 0 | 73.3 | 80 | 89.5 |
3 paths: conclusions about channels need hundreds of paths, this set is an illustration of the method.
Attribution divides credit over the seen touches but does not prove causality: a channel may not have influenced the order. Check decisions with a test against a control.
After restrictions on ad systems the data is incomplete: some touches are outside the digital trace (offline advertising, personal recommendations). Compare channels over the same observation windows.
What to do with the result: a channel that wins in all models is reliable; a channel whose assessment jumps should be checked by an experiment. Do not cut the “first touch” budget by one model.
A typical mistake: taking attribution for causality and comparing channels over different observation windows.
Nuances and limits
- Taking attribution for causality.
- Comparing channels over different observation windows.
- Forgetting that some touches are outside the digital trace.