Somebody opens an email. Twenty nine days later they buy something. Your platform credits the purchase to email, the number goes in the report, and everybody nods (if you're B2C). If you're B2B well, who knows!? They clicked a link in the email, but they also saw us on social? 

Might not have been email at all. Could have been a LinkedIn post they saw that morning, a conversation with a colleague, a podcast, a price change, or the fact that they finally had budget. The email happened to have landed inside the window, so the window claimed it.

Can you ever be sure? No. Not with any model, not with any platform, not at any budget.

And I would argue that is the beauty of email rather than the flaw in it. Email is not a linear channel and it was never going to submit to a linear model. Its effect is compound, most of it happens offline, and the important part happens somewhere no analytics tool has ever been able to reach, which is inside somebody's head.

The models

Why every attribution model undercounts email

Worth going through these properly, because people tend to assume there is a correct model they have not found yet. There is not. Each one fails email in a specific and predictable way.

Last click, or last touch

Email either gets everything or nothing, depending on where it happened to sit. Somebody reads your email, thinks about it for two days, then types your brand name into Google and buys. Branded search takes the credit. Email did the work and search collected the payment.

First click, or first touch

Credits the discovery and ignores everything after it. Email almost never gets first touch, because to be on your list somebody has to have encountered you already. So the model structurally excludes email from the one position it awards credit to.

Linear, or even weighting

Every touchpoint counts the same, which is obviously untrue. An email that changed somebody's mind and a display impression they never consciously saw are not equivalent, and pretending they are is not neutrality, it is giving up.

Time decay

Weights recent touches more heavily, which punishes exactly the work email does. Awareness, education and trust are built over months, so the further back the useful work happened, the less credit it receives. The model is designed to discount compounding.

Position based, usually forty twenty forty

First touch forty percent, last touch forty percent, everything in the middle shares twenty. Ask anybody where those numbers came from and nobody can tell you, because they were chosen because they looked balanced. And email lives in the middle, sharing a fifth of the credit with everything else.

Data-driven and algorithmic models

The most sophisticated and, for email, not much better. A model can only weight what it can observe, and the most important email interaction is frequently the one that produced no event at all. Somebody read the preview text, took the point, and never opened. That is a touch. It is invisible.

And the engagement window model, which most ESPs use

Opened or clicked within a set period, then purchased, so the purchase gets credited. The trouble is that this is correlation wearing a badge. Somebody who opened your email is by definition more engaged with your brand than somebody who did not, and more engaged people buy more often anyway. The model is measuring engagement and calling it causation.

Which is why attributed email revenue so often looks impressive and so rarely survives a proper test.

Free Email Marketing Health Check

Audit your entire ecosystem in under 30 minutes

Answer a set of straightforward questions across every area of email marketing and get a personalised score, traffic-light priorities, and clear actions and improvements you can make today.

Take the Health Check →

Brain attribution

The impact of email happens somewhere you cannot instrument

Now the part I care about, and the reason I have made peace with all of the above.

When somebody receives an email that lands, a series of things happen and almost none of them produce a trackable event.

• They register it. Even without opening. The sender name, the subject, the fact that you turned up again.

• They remember it. Sometimes for months, and usually without being able to say where they heard the thing they now believe.

• They mention it to somebody. In a meeting, on a call, to a colleague at lunch. Your argument, in somebody else's mouth, with no link attached.

• They forward it. To the person who owns the budget, or the one with the problem. You never see that forward and the recipient is not on your list.

• They screenshot it. Into a group chat, a Slack channel, a WhatsApp thread. Entirely invisible to you.

• They think about it. Which is the whole objective of most marketing and produces no data of any kind.

• They add it to a list. A note, a to-do, a tab left open for three weeks. Action deferred, intent formed, nothing recorded.

All of that is attribution. It is simply brain attribution, and there is no pixel for it.

Which means the click is not the effect of your email, it is the small visible tip of a mostly invisible process. Your models are not slightly wrong at the edges. They are blind to the main mechanism and precise about the rest, which is a dangerous combination because precision looks like accuracy.

The only question

If you stopped doing it, what would happen?

That is attribution. Everything else is bookkeeping.

Not which touchpoint deserves the credit, which is an accounting argument. The question is what this channel is contributing that would not happen without it, and there is exactly one reliable way to find out.

Holdout testing, which almost nobody runs

Take a matched group of people who look like the rest of your list, stop sending to them, and measure what they do compared with everybody else. The difference is email's contribution. Not attributed, not modelled, not inferred from a window. Measured.

• Match the group properly. Same engagement profile, same lifecycle stage, same acquisition sources, same value bands. A random split on a list with wildly different segments gives you noise.

• Make it big enough to read. Small holdouts produce differences you cannot distinguish from normal variation, which is how people conclude email does nothing.

• Run it long enough to see the compounding. A two week holdout measures almost nothing, because the mechanism works over months. Think one quarter minimum, two if you can bear it.

• Measure the outcome, not the engagement. Revenue, pipeline, retention, repeat rate. The holdout group cannot open emails, so comparing engagement is meaningless by construction.

• Keep everything else constant. No campaign changes, no pricing changes, no new acquisition push during the window, or you are measuring something else.

• Report the uplift as a range, not a number. Because it is an estimate with a confidence interval, and presenting it as a precise figure invites exactly the false certainty you were trying to escape.

The reason almost nobody does this is that it requires deliberately not sending to a group of people who might have bought, which feels like switching off revenue. It is the only test that tells you the truth, and the discomfort is the price of knowing.

The model

Three layers, reported separately, never blended

If you have to put a model in a report, and most of us do, build it in layers and label each one for what it is.

The leading indicators

Since the lagging one is unmeasurable, use the ones that are not

• Conversion to trust. How many people trusted you enough to hand over inbox access this month, through which routes, at what strength. The first commercial event in your funnel and nobody counts it.

• Engaged list size. The figure that tracks revenue, as opposed to total list size, which tracks your acquisition budget.

• Inbox reach and impressions. How many real people you reached and how many times you landed in front of somebody. Awareness metrics, for an awareness channel, exactly as every other channel has had for years.

• Reply rate. The least gameable engagement signal you have and the one most programmes have designed out of existence with a noreply address.

• Self-reported attribution. How did you hear about us, asked at the point of purchase. Routinely dismissed as unreliable, and it frequently beats the tracked version, because a human telling you what influenced them is a better source than a cookie guessing.

Making peace with it

You have to be all right with not being able to track it fully

The position I would encourage you to take, and it takes some nerve in a business that likes dashboards.

Every brand channel in marketing accepted this decades ago. Nobody asks a billboard to prove itself impression by impression. Nobody demands that sponsorship produce a click path. Nobody cancels a brand campaign because this month's reach did not convert. We understand perfectly well that some marketing works in ways we cannot trace, and we fund it anyway.

Email is subjected to a standard nothing else has to meet, and the reason is an accident of history. Email was measurable early, so the industry built its entire evaluation on the things it could count, and then kept the standard long after the counting broke.

We did not measure what mattered. We made what we could measure matter.

So be all right with it. Run the holdout, report the layers, watch the leading indicators, and stop apologising for a channel whose main effect happens somewhere your analytics were never going to reach.
 

The conclusion

Email attribution is garbage because email is not the kind of thing attribution models were built for. It is compound, it is non-linear, it works on people who never click, and a great deal of its effect happens in conversations, meetings and memories that no platform has ever seen.

The answer is not a better model. The answer is a better question, and the question is what would happen if we stopped.

Run that test once, properly, and you will never argue about last-click again.

Before you measure it, make sure it is arriving

Any attribution work is built on the assumption that your email reached somebody, and around fifteen percent of a typical send does not. A holdout test where the control group and the test group are both partly in spam is not a test.

 

My free 60 minute Email Deliverability Training covers placement, reputation and the metrics worth tracking. The certification programme takes you through the full audit.

Join the free deliverability training →

Further reading from The Vault:

• Email Is an Awareness Channel: Reach, Impressions and the Conversion to Trust

• Conversion to Trust: The Email Metric Nobody Is Measuring

• The Future of Email Marketing

• Inbox Psychology: Why People Stop Seeing Your Emails

• Email Marketing Best Practice Is Mostly Wrong

• How to Know Your Sender Reputation