Les articles d'aide existent en allemand et en anglais. Vous les lisez ici en anglais.
Attribution
How does the data-driven model calculate?
It calculates the removal effect: for each channel it calculates how much less often a path leads to a conversion within the model once that channel is taken out of your paths; normalised to sum to one, that gives its share. Paths without a conversion count too, and they are exactly what sets this apart from a fixed rule. It remains an attribution: whether purchases would have failed to happen without the channel, only an experiment can tell.
What method is behind it, what does it require, and where does it end?
The method in four steps
- 01All paths in a period form a chain: from the start through the channels to a conversion or to a drop-off. A path ends at an order or enquiry; whatever comes after it, such as the visit from the shipping email, starts a new path and gets nothing for that order. Goal steps such as a newsletter sign-up do not end a path. Clicks, visits and, with post-view switched on, viewable impressions count. Consecutive contacts with the same channel count as one.
- 02From the chain follows how likely a path is to convert. That is the baseline.
- 03Then one channel is removed: every path that ran through it ends without a conversion. The probability is computed again. The relative drop is that channel’s removal effect.
- 04The removal effects of all channels are normalised to sum to one. Those are the shares.
An example you can recompute
One hundred paths, eighty of them converting: 60 with search only and a conversion, 20 with display then search and a conversion, 20 with display only and no conversion.
| Channel | Probability without it | Removal effect | Share |
|---|---|---|---|
| Baseline | 0.80 | – | – |
| Search removed | 0.00 | 100% | 80% |
| Display removed | 0.60 | 25% | 20% |
Linear would give display 12.5 per cent: half of each of the twenty converting two-contact paths, so ten out of eighty. The data-driven model gives it twenty per cent, because in the non-converting paths it stands alone and therefore differentiates. A test checks these same figures against the same function that computes your data.
What the model requires
- At least 100 paths, 20 conversions and two channels within 90 days. That is an unlock threshold, not proof that the figure is meaningful. Just above it, the number moves noticeably from day to day.
- The basis is refitted once a day. Cross the threshold today and you see the model tomorrow, not within the hour.
- There is no confidence interval. Publishing a range we do not measure would be invented precision; to judge stability, compare two periods.
Where it reaches its limit
- If every path runs in strict order (A first, then B, nothing else), removing any link breaks the chain. Both channels get the same share and the result equals “linear”. Correctly computed, and still no insight.
- The removal effect describes how strongly a channel differentiates in the observed paths, not how much revenue it caused. Whether the purchases would have failed to happen without it is something only an experiment can say.
- Channels without a measurable contact do not appear. What was not measured cannot be taken into account by this model either.
Note
Why Markov and not Shapley: the Shapley value would need 2^n coalitions, which with twenty channels is over a million per report. Above all, in its usual formulation it only knows channels from converting paths; paths without a conversion would not enter at all, and half the benefit would be lost.
Dernière vérification le 17 oct. 2026
Dans le portail client
Également demandé
- What is the removal effect?
- Which algorithm is behind data-driven?
- Why do I need 100 paths?
- Is data-driven more accurate?
- Do you use Shapley values?
- Does a visit after the purchase count in the data-driven model?
Articles connexes
- How do the attribution models calculate?
Which models exist and how do they distribute the share?
There are seven models: six distribute a conversion by a fixed, disclosed rule, and the seventh reads the rule from your own paths. If a model finds matching touchpoints, it distributes the conversion’s full value across them; if it finds none, the conversion stays “unattributed” and is shown separately. Attributed and unattributed together add up to all conversions it calculates.
- How do I read the model comparison?
Seven models distribute the same conversions differently. How do I read the comparison?
The page compares two models on the same conversions: per channel, the share under model A (your default model) and under model B, plus the change in percentage points. The differences come from the distribution rule and are no evidence of additional advertising effect.
- How do I measure what the advertising actually achieved?
Can Metrunio measure the effect of advertising, not just assign it?
With an experiment: the switchback trial is available from Growth, two campaign groups from Scale and on the agency plan. An experiment compares two groups that should differ in exactly one thing and measures the difference along with its uncertainty. Attribution does not answer this: it says which touchpoint a conversion is credited to, not whether it would have happened without the advertising.