Email Attribution Models: First Touch to Multi-Touch
Five models, the very different pictures they paint of the same campaign, and how to pick one and stop arguing.
On this page
The short version
- Five models will give five different answers about the same campaign, and all five are arithmetically correct. The model is a choice about what you want to credit, not a measurement.
- Last-click systematically under-credits email in long cycles and over-credits it in short ones, which is why the same channel looks strong in retail and weak in B2B.
- Pick one, write down why, and stop relitigating it. Switching models to make a number look better is the failure this whole area invites.
Attribution is the question of which touch gets the credit when several preceded a purchase. It has no correct answer, because the counterfactual — what would have happened without any one of them — is unobservable.
What it has instead is a set of conventions, each of which encodes a different belief about how buying works. Understanding which belief you are adopting is more useful than looking for the accurate model, which does not exist.
The five models on one journey
| Model | Ad | Search | Direct | Believes | |
|---|---|---|---|---|---|
| First touch | £200 | — | — | — | Discovery is what matters |
| Last touch | — | — | — | £200 | The final nudge closes it |
| Last non-direct | — | — | £200 | — | Direct is not a channel |
| Linear | £50 | £50 | £50 | £50 | Every touch contributed equally |
| Time decay | £20 | £40 | £60 | £80 | Recency correlates with influence |
Email goes from nothing to the whole amount depending on the row. Nobody miscalculated — the models are answering different questions.
Why email is the channel most affected
Email usually sits in the middle of a journey. It rarely discovers a customer — they were already on your list — and it often does not close them either, because they go away and come back directly.
That position means first-touch and last-touch both under-credit it, and last-non-direct over-credits it relative to those. The channel's apparent value can double or halve on a decision nobody in marketing made deliberately.
This is also why published claims about email's return vary so wildly. They are not measuring the same thing.
Choosing one you can defend
Last non-direct is a reasonable default for most programmes. It is simple, it is available in every analytics tool without configuration, and it does not discard email in favour of a direct visit that only happened because someone remembered you.
Move to time decay when your cycle is long enough that a single touch clearly is not the story, and you have the discipline to keep the window consistent.
Position-based models — crediting the first and last touch heavily — suit programmes where acquisition and closing are genuinely distinct efforts. They are harder to explain, which matters more than it sounds when the numbers reach someone outside marketing.
The holdout, which settles arguments models cannot
The only way to know what email contributed is to not send it to some people. A randomly selected holdout that receives nothing, compared against the rest over a quarter, gives you an incremental figure that no attribution model can produce.
It costs whatever those people would have bought, which is why it is usually a small percentage and why it is worth running on the flows rather than the campaigns — an automated sequence runs long enough for the comparison to accumulate.
A single holdout result is worth more than a year of arguing about models, and almost nobody runs one.
Attribution check
- One model is named as the number of record
- The reason it was chosen is written down
- The lookback window is stated and consistent
- Everyone reading the report knows which model produced it
- A holdout exists on at least one automated flow
- Nobody has proposed changing the model since the last bad quarter
Frequently asked questions
Which model makes email look best?
Last non-direct, usually, which is exactly why choosing on that basis is a mistake. If you pick it, pick it because direct visits are not really a channel, and be able to say so.
What lookback window should I use?
Longer than your typical time from first touch to purchase, which you can measure. A thirty-day window on a ninety-day cycle discards most of what you are trying to credit.
How big should a holdout be?
Large enough that the difference resolves, which is the same sample size question as any test. On a small list this may mean running it for a quarter rather than a fortnight.