Your attribution model is a policy, not a measurement

Three dashboards, one quarter, three different numbers for the same channel. Paid search reads £186k in the ad platform, £94k in the CRM report, and £121k in the board pack. Nobody has made an error. The three systems are running three different attribution models, and each is doing precisely what it was told.

The instinct is to go and find the right number. There isn't one. A revenue attribution model does not measure what caused the revenue — it is a rule you chose for dividing credit between touches that all actually happened. Once that lands, the question stops being which model is correct and becomes which rule produces a decision I am willing to defend in front of the person whose budget it moves. That second question has an answer.

Splitting credit is an accounting choice, not a discovery

A deal closes at £40,000. Along the way the buyer clicked a LinkedIn ad, read two blog posts from organic search, ignored three emails, attended a webinar, and finally arrived through a branded search before booking a call.

Every one of those things happened. The revenue is a single indivisible £40,000. Attribution is the act of cutting that £40,000 into pieces and handing them to touchpoints, and there is no fact of the matter about where the knife goes. The counterfactual — what this buyer would have done without the webinar — is unobservable. You cannot run the quarter again.

So the model is a policy. Policies get judged on whether they lead to good decisions, not on whether they are true.

The models you will actually be offered

W-shaped is position-based with a fourth anchor at opportunity creation, which matters if your CRM has a stage worth anchoring to. If it doesn't, adding the anchor invents precision you haven't earned.

Every one of these is arithmetic. You can implement any of them in SQL over a touch table in an afternoon. That is worth saying plainly, because the tooling market prices them as if they were something harder.

"Data-driven" is a different kind of claim, and it needs volume

Data-driven attribution — Google's DDA, HubSpot's equivalent, the Shapley-value approach most vendors have converged on — is genuinely different. Rather than applying a fixed rule, it compares converting and non-converting paths and assigns credit to the touches that shift conversion probability. That is a real statistical claim, not an accounting convention.

It also has a real requirement: enough conversions, across enough distinct path shapes, for the differences to mean anything. Google has historically wanted conversions in the hundreds per month before DDA is offered on a conversion action. Most B2B companies doing forty deals a quarter are nowhere near it.

Below that threshold, data-driven attribution does not fail loudly. It produces confident-looking percentages built on a handful of paths, and those percentages move substantially month to month for no reason connected to your marketing. A rule that is transparently arbitrary is safer than a model that is opaquely noisy — at least everyone knows where the arbitrariness is.

The model is answering a narrower question than the one you asked

Most attribution arguments are really scope arguments. Three different questions get asked in the same meeting:

• Where should next quarter's budget go? This is about marginal return — what happens if I add £10k to this channel. Attribution answers it badly under any model, because splitting historic credit says nothing about the next pound. Incrementality testing and geo holdouts answer it properly.
• Which channels bring us people we would never otherwise meet? First-touch, unapologetically. That is the one thing it is good at.
• Which activity is efficient at converting people already in-market? Last-touch or time-decay, and you accept that branded search will look like a hero.

None of those three is wrong. They answer different questions, and the £92k spread between the extremes is not an error to be reconciled — it is the finding. A channel whose number swings that far between first and last touch is doing one job well and the other badly, and knowing which is more useful than a single blended figure.

Run two on purpose, and report the gap

The practice that survives contact with a real business: pick two models that sit at opposite ends, report both every month, and treat the difference as its own metric.

First-touch and last-touch is the usual pair. A channel that scores high on both is genuinely carrying the pipeline. High first, low last means it opens doors it cannot close — likely under-credited by whatever single model you were using before, and probably under-funded. Low first, high last usually means the channel is harvesting demand created somewhere else, which is fine and cheap right up until you cut the thing that was creating it.

This is harder to put in a board pack than one number. It is also the only version that does not quietly mislead. If you are building the reporting layer for this, the design decisions that make a dashboard trustworthy matter more than the model choice — a single unlabelled figure invites exactly the argument you are trying to end.

The join breaks before the maths does

Model selection is the interesting part of the conversation and almost never the reason the numbers are wrong.

The reason the numbers are wrong is that the acquisition source did not survive the trip from the click to the closed-won deal. The utm_source was captured on the landing page and written to a form field. The form created a contact. Somebody later merged that contact with a duplicate, and the merge kept the other record's blank source. Or the deal was created manually by a rep from a conversation, with no link back to the contact at all. Or the source field is free text and contains linkedin, LinkedIn, li, and Linkedin ads as four distinct values.

Before you argue about time-decay half-lives, run this: take last quarter's closed-won deals and count how many can be traced to a first touch at all. If it is 60%, your model is being applied to 60% of the revenue and silently ignoring the rest — and the missing 40% is not random, because manually-created deals and merged records skew heavily towards sales-led and referral paths. Fixing the data quality underneath the join moves the number more than any model change will.

Write the rule down where anyone can find it

Whatever you pick, the failure mode is not choosing badly. It is that six months later nobody can say which rule is in force, the ad platform is quietly applying its own, and two people are comparing figures produced by different policies without knowing it.

So: one sentence, in the same place as the metric definitions. Revenue is credited to the first known non-direct touch on the associated contact, using the deal's closed-won date, excluding deals with no linked contact. Boring, specific, and it ends the argument before it starts.

The check worth doing today

Open the report your board sees and ask two questions of it. Which model produced this number? And what percentage of closed revenue does it cover?

If the first has no answer, you do not have an attribution model — you have whatever default your tooling shipped with. If the second is well under 100%, you have a join problem, and no amount of model sophistication will fix a percentage of revenue the model never sees. Both are more useful to know than which of the six rules is best.