How to measure channels that do not attribute cleanly

Why can't better tracking fix this?

Because the connection you want to make does not exist in any system. Attribution works by following an identifier from an impression to a conversion. When the impression is audio, printed, spoken by a colleague, or seen on a device the person never buys from, there is no identifier to follow. This is a property of the situation, not a gap in your setup.

Consent rules, cross-device behaviour and in-app browsers make the tracked portion smaller than most dashboards imply, but even with perfect consent and one device per person, the podcast listener and the conference attendee remain invisible. Vendors selling attribution products tend to describe this as a solvable problem because their product only works if it is one.

The consequence is that your analytics is a report on the channels that happen to be trackable, not a report on your marketing. Treating it as the latter systematically defunds everything else, which is exactly how companies end up spending their entire budget on the last click before the sale.

What is last-click actually telling you?

Which channel was present at the end, which is close to useless for deciding where to spend. Branded search is almost always the last click, because people search your name before buying, so a last-click report will reliably tell you that the cheapest way to get customers is to advertise to people who already decided to become customers.

This produces a specific and common failure: the report says brand search has an excellent cost per acquisition, budget shifts towards it, and the channels that created the brand search in the first place get cut. Six months later the volume of branded search falls and nobody can explain why, because the thing that caused it was cancelled two quarters earlier and the effect was never attributed to it.

A quick diagnostic: look at what proportion of your reported conversions come from branded search and direct. If it is the majority, your attribution model is mostly describing demand you already created, and you have very little information about what creates it.

Which methods work when tracking does not?

Four, and they trade precision for validity. None of them will tell you which impression caused which sale. All of them can tell you whether a channel is doing anything, which is the question that actually decides the budget.

How do you run a holdout test without a data team?

Turn the channel off, everywhere, for a defined period, and watch total enquiries rather than that channel's own numbers. This is the cheapest causal test available and most companies never run it because switching off feels reckless. The information it produces is better than anything a dashboard will give you.

The window has to be longer than your sales cycle plus the memory effect, which is why a two-week blackout on a channel with a three-month consideration period proves nothing. Pick a period with no seasonal distortion, no product launch and no competitor event you know about, and write down beforehand what change would count as a real effect.

A softer version is the geographic split: keep the channel running in one region and pause it in a comparable one. This controls for seasonality and news, which is the main weakness of a straight blackout, at the cost of needing enough volume in each region for the difference to be readable.

Is asking customers how they found you worth doing?

Yes, with a clear understanding of its bias. Self-reported attribution systematically over-credits memorable channels like recommendation and events, and under-credits advertising, because people do not remember or do not like admitting that an ad worked. It is directionally useful and numerically unreliable.

It becomes considerably more useful when you change the question. Instead of asking how they heard about you, ask what made them get in touch now, and leave it as a free text box. The answers tell you about triggers and timing, which are harder to get any other way and more actionable than a channel name.

Ask it at the point of enquiry rather than after purchase, and ask it of everyone rather than sampling. It costs one field on a form, and after a few hundred responses the patterns in the free text are usually clearer than the channel report.

What should you accept you will never know?

The contribution of any individual impression, and the exact split between channels in a multi-touch journey. Accepting this is not defeatism, it is the precondition for measuring anything useful, because effort spent chasing that precision is effort not spent on the tests that would actually change a decision.

What you can know, reliably: whether total demand moves when a channel starts or stops, whether branded search grows, whether the cost of your captured demand is rising, and what customers say triggered the enquiry. Those four are enough to run a budget.

A working discipline is to hold each channel to the measurement it can support and refuse to hold it to any other. Direct response channels answer to cost per acquisition. Demand creation channels answer to holdout tests and branded search trends. Mixing the two produces a report that looks rigorous and is quietly wrong.