Attribution Records Where the Click Happened, Not Where the Decision Was Made

Your attribution report is probably correct. That is the problem. It accurately records that someone searched your company name and clicked a paid ad, and it credits that ad with the sale. What it does not record is that the person decided to work with you six weeks earlier, after a colleague mentioned you in a Slack channel, and that the ad simply collected a decision that had already been made. Both statements can be true at once. The model is right about the click and wrong about the cause, and there is no configuration that fixes it, because the missing information was never in the system.

The two things being confused

An attribution model observes trackable interactions on properties you control. It builds a path from those observations and distributes credit across it. Within that scope it works. The failure is a category error: the path is a record of arrival mechanisms, and a purchase decision is a change of mind. Those overlap only sometimes.

The change of mind usually happens somewhere unobservable. Someone recommends you in a private community. A prospect watches half a conference talk. A former colleague who used your product joins a new company and brings the name with them. Your name comes up in an AI assistant's answer to a research question with no referrer attached. None of these produce a session, a UTM parameter, or a row in any table, and all of them regularly do the persuading.

What the model sees afterwards is the arrival. The person, now convinced, types your name into a search engine or clicks a retargeting ad they have seen forty times. The model records that final step and hands it the credit. This is why branded paid search and retargeting so reliably report extraordinary returns: they sit closest to the moment of arrival and furthest from the moment of persuasion.

Why changing the model does not help

The standard response is to move from last click to something more sophisticated. It is worth understanding why each option lands in the same place.

Model	What it assumes	Where it breaks
Last click	The final touch caused the sale	Credits collection channels, starves discovery
First click	The initial touch caused the sale	Ignores everything that sustained interest over months
Linear	Every touch mattered equally	Nothing about buying works this way
Time decay	Recent touches mattered more	Encodes the last-click bias with extra steps
Data-driven	Patterns in observed paths reveal contribution	Only reasons over the touches it can see
Multi-touch with CRM	Sales-logged touches fill the gaps	Depends on what a busy salesperson wrote down

Every row shares one assumption: that the influential interactions are present in the dataset. When the persuasion happened off-platform, all models redistribute credit among the wrong set of touches, and they do so with increasing statistical confidence as they get more sophisticated. A data-driven model applied to a path that omits the cause is precise about the wrong thing. Our guide on what marketing attribution cannot tell you works through this boundary in detail, because knowing where the instrument stops reading is more useful than buying a better instrument.

The practical damage is budget reallocation. Attribution reports are used to move money from channels with weak attributed return to channels with strong attributed return. That movement systematically transfers budget from where decisions are made to where clicks are collected. Run it for four quarters and you end up with a marketing function that is very efficient at harvesting demand and no longer creates any.

What to do instead of arguing about models

The useful posture is to treat attribution as one instrument among several, with a known blind spot, and to cover that spot with methods that do not depend on tracking.

1. Ask buyers directly. A single open-text field on the enquiry form asking how they first heard of you, with no dropdown, produces messier and more truthful data than any model. Read the answers rather than categorising them.
2. Watch branded search volume as the summary signal. It moves when people remember you, regardless of which untracked channel caused the remembering, which makes it the best available proxy for the invisible half of the funnel. Our guide on brand search volume as a marketing outcome covers how to separate navigational from evaluative queries so the number stays meaningful.
3. Run holdouts where you can. Turning a channel off in one region or segment for a defined period tells you what it contributes far more reliably than any modelled credit, and it is the only method in this list that establishes causation rather than correlation.
4. Measure the untrackable channels on their own terms. A podcast, a community and a conference talk each have leading indicators that are not sessions. The approach in our guide to measuring channels that do not attribute sets those out without inventing a tracking story that does not exist.
5. Report the unknown as a line item. If forty percent of new customers cannot be traced to a source, put that on the slide as a figure rather than distributing it across the channels you can see. Visible uncertainty produces better decisions than hidden uncertainty.

The organisational part

This is not only a measurement question, and it is worth being clear about that. Attribution reporting is often the mechanism by which marketing justifies its existence internally, which means anything that reduces its apparent precision is threatening. A team that reports a confident attributed return is more comfortable than a team that reports a range and an honest unknown, even when the second is more accurate.

That pressure is strongest in organisations early in building a marketing function, where the person or agency running it is also the person presenting the numbers. The incentive to report clean attribution is highest exactly where the underlying data is weakest. It is one of several considerations in the decision between a first marketing hire and a first agency, because who owns the reporting shapes what the reporting says.

The fix is to agree in advance which decisions attribution is allowed to make. Optimising ad creative and bidding within a channel: yes, the data supports that. Deciding whether a whole channel deserves to exist: no, the data cannot see far enough. Setting that boundary before the next planning cycle prevents the argument happening under budget pressure, when it is always resolved in favour of whichever number is easiest to defend.

What to do next

Take your last twelve closed deals, ignore the CRM source field, and ask the salesperson who handled each one how the buyer first came across you. Write the answers down next to what your attribution report says about the same deals. The gap between the two lists is the size of your blind spot, and it takes an afternoon to measure.

If that gap turns out to be large, the next question is what to do about the channels you have been underfunding, which is the kind of allocation problem our digital marketing practice is usually brought in to work through.