Why you cannot tell which channel drove the sale

Consider an ordinary path to a purchase. Someone sees a post on social media in March and does nothing. In April they search your category and click a paid advertisement, look around, and leave. In May a friend mentions you. In June they search your business by name, click the unpaid result, and buy. Which channel produced that sale?

Every available answer is defensible and none is correct. Last-click attribution credits organic search, which arguably did nothing except be present at the end — the buyer was already looking for you by name. First-click credits social, which started everything and could not have closed it. A multi-touch model splits the credit across the touchpoints it can see, using a rule someone chose. And the friend's recommendation, quite possibly the decisive event, is invisible to every model because it left no trace in any system. The disagreement between models is not a technical problem awaiting a better tool. Causation in a multi-step journey is genuinely underdetermined by the available data, and each model is an opinion about how to divide credit, expressed as arithmetic.

Last-click is the default in many tools because it is simple and unambiguous, and it systematically overcredits whatever people do immediately before buying — typically branded search and direct visits, which are the channels that capture existing intent rather than create it. Optimise on last-click alone and you will reliably conclude that the activities creating demand are not working, and cut them, and then watch branded search decline some months later without an obvious cause.

First-click has the mirror bias: it overcredits discovery and ignores everything that turned interest into a decision. Linear models split credit evenly, which assumes every touchpoint mattered equally — a straightforward assumption and almost certainly false. Time-decay models weight recent touches more heavily. Data-driven models, including GA4's, use observed patterns to distribute credit, which is more sophisticated and still a model: it is inferring credit from correlations in incomplete data, and its workings are not something you can inspect line by line. The important property all of these share is that they are internally consistent and mutually contradictory, and no data in your possession adjudicates between them.

Even a perfect model would be working from a partial record. Word of mouth leaves no trace, and for many small businesses it is the largest single source of customers. Offline conversations, a shop sign, a vehicle livery, a mention in a group chat, a printed leaflet, someone recognising your name from a previous job: none of it is measurable, all of it works. Cross-device journeys fragment one person into several. Consent banners and blockers remove visitors from the record entirely. Long consideration periods exceed the windows within which tools connect touches.

So the measured portion of the journey is a subset of the real one, and the subset is biased rather than random: it systematically over-represents whatever happens online, recently, and on one device. This means the confident channel report in any analytics tool describes the part of reality the tool can see, presented as though it were reality. Recognising that is not defeatism. It changes what you ask the data for — direction and rough magnitude rather than precise credit — and it stops you from making irreversible decisions on the basis of a decimal place that was never measured.

The honest and useful alternative is to stop trying to attribute individual sales and start watching aggregates that do not require it. Blended acquisition cost is the primary one: everything spent on getting customers in a period, divided by all new customers gained in that period. It requires no attribution model, no assumptions about credit, and it is arithmetic rather than inference. Track it monthly, and its direction tells you whether your marketing overall is becoming more or less efficient — which is the question that actually governs your spending.

Alongside it, track spend and volume per channel separately without claiming causation, and use the crude test that survives all of this: turn something off. If you stop a channel and total new customers fall over the following period, it was contributing something. If nothing happens, it was not, whatever the attribution report claimed. This is genuinely informative because it operates on the aggregate you care about rather than on a modelled credit assignment. It is imperfect — other things change at the same time, and delayed effects can take months to appear — so do it deliberately, one channel at a time, over a long enough window, and expect a weak signal rather than proof.

The most underused source of attribution information for a small business is asking. A single optional question at the point of purchase or enquiry — how did you hear about us, with a short list and an open field — produces something no analytics tool can: the customer's own account of what brought them. It captures word of mouth, offline exposure and the things that happened outside any measurable channel, which is precisely the portion the models cannot reach.

It is also unreliable, and the unreliability should be stated rather than hidden. People misremember, they name the most recent thing, they cannot distinguish paid search from organic and have no reason to, and the ones who answer differ from the ones who do not. What makes it valuable despite that is that its errors are different from the analytics tool's errors, so where self-reported answers and measured channels agree you have genuine corroboration, and where they disagree you have learned that something is happening outside the measured record. Two imperfect and independent views beat one confident and incomplete one, which is the general principle worth taking from all of this.

The conclusion of an honest attribution exercise is often that you cannot tell which channel drove a particular sale, and that this is a property of the situation rather than a shortcoming to be apologised for. The valuable skill is being precise about the uncertainty rather than resolving it falsely. "Organic search was the last touch on most sales, our blended acquisition cost is stable, and we cannot separate the contribution of social from word of mouth" is a genuinely useful position. A single number claiming a channel produced a specific share of revenue is a modelling artefact presented as a measurement.

The practical discipline is to label figures with how they were produced. A count of orders is a count. Total spend divided by new customers is arithmetic. A per-channel revenue figure is a model output, and naming which model it came from is what lets the next person judge it. Where you must make a decision without adequate evidence — which is often — make it, and write down what you assumed and what would change your mind. That leaves something to revisit, which is more than a false precision leaves. And keep the reversible decisions reversible, because the honest state of attribution knowledge for most small businesses does not support betting the business on a channel report.