Attribution is mostly a story you tell yourself

A marketing lead once told me their attribution dashboard had assigned forty-one percent of a quarter's revenue to organic search. The number had a decimal point. It looked like the output of measurement, the same way a scale looks like the output of measurement when you step on it. Nobody in the room asked what the model would have said if they'd built it with a different weighting scheme, because nobody in the room had built the model — a vendor had shipped it with a default, and the default had never been questioned. That number was not measured. It was assumed into existence by a rule someone chose before any of the quarter's deals had happened, and it would have produced a different decimal point under a different rule with the exact same customers doing the exact same things.

That is not a complaint about the tool. It is a claim about what attribution is, and it is worth being precise about it before going further, because the honest version of this argument is more useful than the cynical one. Attribution is not lying to you. It is answering a question you did not ask — "how would this rule split credit across the touches we happened to observe" — while presenting the answer as if it were the question you did ask, which was "what caused this sale." Those are different questions, and the gap between them does not close with a better model. It closes with more data than most companies will ever have, or it does not close at all.

The models are rules, not findings

Say what each one actually does, plainly, because the marketing language around attribution obscures this.

First-touch gives all the credit to whatever channel introduced the customer, usually the first recorded ad click or organic visit in an analytics tool's cookie window. Last-touch gives all the credit to whatever channel preceded the conversion event, often literally the final click before a form submit. Linear splits credit evenly across every touch the tooling managed to record. Time-decay weights recent touches more heavily. Position-based splits credit between the first and last touch with a smaller share in the middle. U-shaped, W-shaped — the naming gets more elaborate, but every variant on this list is a rule for dividing a fixed pie, chosen in advance, applied afterward. None of them was derived from evidence that customers actually respond in the pattern the rule assumes. They were chosen because a rule has to be chosen, and these are the rules that are easy to compute from the data a browser happens to expose.

This matters because "model" is doing rhetorical work it hasn't earned. Call something a model and it sounds discovered — recovered from the data through some process resembling science. Call it what it is — an allocation convention — and the forty-one percent stops looking like a finding and starts looking like a choice that could have gone differently. Run the same quarter's data through last-touch and then through linear and watch the split move by double digits with not one additional fact entering the calculation. That volatility is not a bug in a particular vendor's implementation. It is what happens whenever you divide credit among things you cannot actually separate, using a rule that was picked for tractability rather than derived from cause.

Below a few hundred deals, the model cannot be told apart from noise

Here is the part that rarely gets said out loud, and it is a stronger claim than "the models disagree." Below a certain volume, the data cannot tell you which model, if any, describes what actually happened — not because your tracking is bad, but because the problem is what statisticians call unidentified: there is not enough independent variation in the paths customers take for any model to be distinguished from any other by evidence alone.

Multi-touch attribution needs customers to take genuinely different paths, in large enough numbers, for the model to have anything to fit against. A company converting fifteen deals a month, most of which touch two or three channels in some combination of organic, a newsletter and a referral, does not have fifteen data points about channel effectiveness. It has fifteen anecdotes, and a handful of possible path shapes repeated so few times that swapping which deal happened to touch which channel first would change the entire output. The same arithmetic that makes A/B testing pointless at low volume — covered at length in measuring what a website does — applies here with additional force, because attribution isn't splitting one outcome into two buckets, it is trying to apportion credit across a combinatorial explosion of possible touch sequences with a fraction of the sample. A company running twenty deals a month across four channels does not have twenty points of evidence about those four channels. It has, at best, a few examples of each of a dozen or more possible sequences, which is not enough to distinguish "channel A drives conversions" from "channel A happened to be present in the three deals that would have closed anyway."

None of this is an argument that larger companies have solved attribution. It is an argument that they have enough volume for the model to be identifiable at all — which is a much lower bar than accurate, and one that most companies reading a monthly report from a marketing platform have not cleared.

The one field that beats every model, for what it actually costs

If the model can't be trusted below volume, the honest move is not to build a better model. It is to ask the customer, cheaply and imperfectly, and treat the answer as directional rather than exact.

The wording matters more than it should: How did you hear about us? — free text, or free text with a handful of suggested options, placed on the enquiry or checkout form itself, optional rather than required. Not a dropdown that forces a single choice from a list someone wrote eighteen months ago and never revisited; not gated behind a second screen that a fraction of people will abandon rather than fill in. One line, asked at the moment someone has already decided to convert, when answering costs them nothing they weren't already prepared to give.

Its biases are real and worth naming rather than hoping nobody notices. People misremember. They collapse a three-step journey into whichever step is easiest to name — the search that found you gets credit over the podcast that made them search in the first place, because the search is what they can recall doing. A meaningful share skip the field entirely, particularly if it isn't optional and they resent being asked. None of that makes any individual answer reliable. What redeems the method is that you are not trying to attribute individual deals — you are trying to find out whether a channel exists in the minds of the people who converted, at all, which is a far weaker and far more answerable question. Ten people typing "a friend recommended you" into a free-text box in one quarter is a fact about your business that no pixel will ever produce, and it costs a single form field rather than a subscription to a tracking platform.

Read it in aggregate, quarterly rather than weekly. At low volume a monthly read is four answers split across three categories, which is not a distribution, it is three data points dressed up as a chart. A quarter gives the categories room to actually separate from noise.

What no field and no model will ever see

Some of the most consequential channels in a small company's growth are structurally invisible to both the pixel and the form field, and it is worth being honest about that rather than implying the self-reported question closes the gap entirely.

A private Slack channel where someone recommends a vendor. A WhatsApp thread. A conversation at a conference that surfaces a company name three weeks later when someone finally searches it. An internal recommendation passed between colleagues who never mention where they first heard it because, by the time they're filling in your form, they've forgotten — this is what the analytics literature calls dark social, and no amount of UTM discipline recovers it, because it never generated a trackable link in the first place. Even the self-reported field only catches it when the person filling in the form happens to remember and happens to write it down, which is a subset of a subset.

The honest position is that some fraction of every company's growth is simply outside the measurement system, permanently, and the size of that fraction is itself unknowable — you cannot measure the thing you cannot see well enough to know how much of it there is. That should change how confidently anyone states a channel split, not just for the fraction attribution gets wrong but for the fraction it cannot see at all.

What to do with the budget decision anyway

None of this means the budget decision goes unmade. It means the decision should stop pretending to rest on a number precise enough to have a decimal point.

Use the self-reported field as a directional check against platform-reported numbers, specifically watching for the gap between them rather than trusting either side alone: a channel the platforms credit heavily that nobody ever mentions unprompted is worth scrutiny before the next renewal; a channel people keep naming that the platforms barely register is probably underfunded relative to its real effect. Track total outcomes over time against total spend, in the plain sense — did the number of qualified enquiries move after a channel started or stopped, at a scale large enough to be visible against the noise floor described above — rather than asking any model to assign fractional credit within a single quarter's deals. For paid search specifically, where the model-versus- budget question shows up most often, the comparison in SEO versus paid search on a small budget is a better starting point than any attribution dashboard, because it reasons from channel economics rather than from disputed credit.

And write the actual sentence in the report: attribution cannot currently tell us how much of this quarter's revenue came from paid social, and building a model that claims otherwise would be less honest than saying so. Pair that sentence with the outcome-counting discipline laid out in conversion rate: what good looks like, and you have replaced a number nobody can defend with a smaller set of numbers everybody in the room can actually stand behind.

The forty-one percent was never wrong, exactly. It was just answering a question nobody in that room had thought to ask, dressed as an answer to the one they had.