Tags: analytics commerce concept
Multi-Touch Attribution
Date: 2026-08-16
Spreading credit across every touchpoint on the path rather than giving it all to one. It’s more realistic than last click and no more causal — and its accuracy is capped by identity, not by the sophistication of the model.
What it is
Multi-touch attribution (MTA) divides a conversion’s credit across the several touchpoints that preceded it, using a rule or a model.
It exists because single-touch attribution is obviously wrong: crediting one interaction for a decision made over three weeks and six touches describes nothing.
The problem it doesn’t solve
MTA is a more sophisticated allocation of the same fixed credit. Every model divides one conversion among the touchpoints it can see. None of them establishes that any touchpoint caused anything — they can’t, because the counterfactual isn’t in the data.
So the improvement over last click is real but narrower than it sounds: you get a less absurd distribution, not a causal answer. Only Incrementality Testing gives you that.
What actually limits it
Ranked by how much damage each does — and note that the model is last on the list.
1. Identity. MTA requires linking touchpoints across sessions, devices and weeks. Every failure collapses the path towards the final session.
true path display → paid social → organic → email → direct
observed (identity lost) ......... organic → email → direct
credited three channels, two of which were incidental
A 30-day path can’t be reconstructed with a 7-day identifier. For a large share of traffic, MTA is quietly operating on the last few days of a longer journey — Identity Stitching, Browser Privacy Restrictions.
2. Visibility. Walled gardens don’t share path data. Impressions on platforms you don’t own are absent unless the platform tells you, and it tells you in its own model — Walled Garden Reporting.
3. Lost referrers. Anything landing in Direct is a hole in the path — Direct Traffic and Lost Referrers.
4. The window. Touchpoints outside it don’t exist to be credited — Attribution Windows.
5. The model. Last, and it’s where the discussion usually starts.
In plain terms: arguing about linear versus time-decay while a third of your paths are truncated is optimising the last 5% of the problem.
The models
Rules-based and data-driven forms, with worked arithmetic, are in Attribution Models — the same £200 order splitting five ways. What matters here is what the choice is for.
Where it earns its keep
Not as a source of truth. As a diagnostic:
- Compare first-touch and last-touch. A channel scoring far higher on first is an opener; higher on last is a closer. The disagreement between models is real information about where in the path a channel sits — and it doesn’t require either model to be correct
- Find channels that never appear alone. A channel present on many converting paths but rarely as the only touch is doing assist work that last click erases entirely
- Size the gap against incrementality. Where you’ve run a holdout, compare its answer to MTA’s for the same channel. The difference is that channel’s over-credit, and it tends to be largest for the bottom-funnel channels MTA flatters — Geo Holdout Tests
Practical position
- Don’t buy MTA to fix attribution. Fix identity first; the model is downstream of it
- Report a range, not a number. “Between the first-click and last-click estimates” is honest and usually more decision-useful than a single figure with false precision
- Use MTA for within-channel decisions — which creative, which campaign, where the same measurement bias applies to both sides of the comparison
- Use incrementality for between-channel budget decisions, where the bias differs by channel and MTA is most misleading
- Never change model and window at once, and restate history when you do change either — Metric Drift