Why Your Best Performing Channel Is Probably Being Miscredited

Somebody in the meeting says paid search is the best channel. It has the lowest cost per acquisition, the cleanest reporting and the most conversions sitting next to its name. Everyone nods. Budget moves towards it. For two quarters the numbers hold, and then they stop holding, and nobody in the room can explain what changed.

The channel attribution error in one sentence: the channel standing closest to the finish line collects the credit for the whole race.

This is not an analyst making a mistake. It is a structural property of measuring anything: what can be observed absorbs the credit owed to what cannot. Fixing it does not require a modelling project or a new platform. It requires one deliberately blinded test and the discipline to read a total rather than a line item, which is the same standard we apply to attribution you can defend when a board asks where the leads came from.

The shape of the channel attribution error

Attribution assigns credit to touchpoints it can observe. That sounds neutral. It is not, because observability is distributed extremely unevenly, and it is distributed in a way that favours the end of the journey.

A click is observable. A conversation in a group chat is not. An ad impression is observable because the platform selling it also reports on it, which is a conflict of interest you have agreed to in writing. A recommendation made over lunch leaves no trace anywhere, and it may have done more work than every impression that month.

So the ledger carries a systematic bias, and it runs one way: towards the bottom of the funnel, and towards channels that report on themselves. If your reporting looks like any of the following, the error is already operating.

• Your best channel is also your cheapest, and has been for longer than seems plausible.
• Branded search converts unusually well and keeps growing without anyone doing anything to it.
• Direct traffic is one of your largest sources and nobody can account for it.
• Every channel report, added together, claims more conversions than the business actually recorded.
• The activity you cut last year seemed fine for a few months, and then something unrelated started getting worse.

That last one is the tell. Demand creation has a lag, and the lag is longer than the reporting period, which means the cost of cutting it always lands in a quarter where somebody else gets blamed.

How measurable channels absorb the unmeasurable ones

The absorption is mechanical, not conspiratorial. A person is persuaded somewhere you cannot see, arrives through something you can, and the something you can see writes it down. Repeat that for a year and the ledger stops describing the business.

Multi-touch models are often proposed as the answer. They help at the margin, and they do not solve this, because a model can only distribute credit among the touchpoints it can see. Widening the window does not make an unobserved conversation appear in the data; it just spreads the same visible credit more politely.

Consent has made this sharper. Under the European data protection regime and the practices that grew around it, a visitor can decline the tracking a model depends on, and their visit still happens. The measurement layer rarely reports that visit as unknown. It reports it as direct, which is a different claim entirely, and a flattering one.

Brand search: paying for a name they already knew

The cleanest example is brand search. Someone hears about you on a podcast, from a colleague, at a conference. Three days later they type your name into a search box. An ad for your own name appears above your own organic result. They click the ad. Paid search books a conversion at a delightful cost per acquisition, and the podcast gets nothing.

Nothing about that is fraud. All of it is misattribution. The demand was created somewhere else and harvested here, and you paid for a position you had already earned. Branded queries are the point at which an organic result and a paid result compete for the same person, and the basic mechanics of how that organic result is produced are set out plainly enough in Google’s own starter guidance.

The test almost nobody runs

Turn brand bidding off, on a schedule, in a controlled way, and watch total demand for your name rather than the paid line. Two outcomes are possible and both are useful.

Either organic absorbs nearly all of the clicks and total conversions from brand terms barely move, in which case you were buying traffic you already had. Or total brand conversions fall, in which case the spend was doing real work: usually defending against a competitor bidding on your name, or a marketplace listing outranking you. The distinction between harvesting and defending is the only thing the test needs to settle.

When brand bidding earns its place

We are not against brand bidding, and this is worth saying clearly because the argument is often taken further than it deserves. It earns its keep when competitors bid on your name, when your organic result is not the first thing on the page, when your name is also an ordinary word, and when you need control over the message during a rebrand, a live offer or a bad news week. In all of those cases, keep it. Just stop counting it as acquisition.

The discovery-to-conversion gap

Between the moment someone first hears of you and the moment they buy, there is a gap. Days for something cheap, months for anything considered. Almost nothing inside that gap is measurable by default, and it is where most of the persuasion actually happens.

A link pasted into a private message arrives with no referrer. A link opened from a document arrives with no referrer. Several mobile apps strip it, some browsers withhold it, and a visitor who declined tracking arrives carrying nothing at all. Every one of those is a real person with a real reason for coming, and your reporting files all of them under the same shrug.

What the questions tell you that the referrer cannot

There is one place where signal is recoverable, and it is the moment a visitor asks something. What they type is first-party evidence of why they came, and it is more honest than any parameter. A widget such as Flidu sits exactly there: website-informed answers alongside the contact and conversion actions, in one place, so the question and the action are recorded together instead of in two systems that never meet.

Read a month of those questions and the working campaign usually identifies itself. Not because anything attributed it, but because people volunteer where they came from, unprompted, in their own words. "Saw your talk about pricing" is a stronger data point than a tracking parameter somebody forgot to set. This matters most when there is no existing search demand to capture at all, which is the harder situation described in marketing a product nobody is searching for yet.

Direct traffic is not a channel

Direct is not a source. It is the folder where measurement failures are filed, and because it converts well it gets read as a success rather than as a gap.

What lands in it: typed addresses, bookmarks, links from messaging apps, links inside documents and PDFs, applications that remove the referrer, secure-to-insecure hops, visits from people who declined tracking, and every mistagged campaign anyone has ever run. Those are not one behaviour. They are eight, and only the first two are what the label claims.

Treat direct as an unlabelled bucket, not a source. The useful question is never how to reduce it. It is which of your unmeasurable activities makes it move, and the answer to that comes from timing, not from tagging.

A cheap test you can run this quarter

You do not need a data science project. You need one blinded experiment and the patience to leave it alone while it runs.

1. Pick the channel you suspect of collecting other people’s credit. Brand search first, retargeting second. Both are the usual beneficiaries because both sit closest to the purchase.
2. Define the metric before you touch anything. Total conversions and total revenue for the whole business. If the metric is the channel’s own reported number, the test is already meaningless.
3. Split by something the platform cannot optimise across. Geography, a random half of an audience list, alternating weeks. Splitting by campaign is not a split, because the system will simply reallocate around it.
4. Switch it off in the held-out half and change nothing else. Not the landing page, not the offer, not the budget elsewhere. One variable, or you will spend the next month arguing about which one moved.
5. Wait longer than is comfortable. Long enough to cover your real consideration window plus the lag between first contact and purchase. Stopping early is the single most common way this test returns a wrong answer.
6. Read the total, never the channel. If the business is flat, the channel was harvesting demand. If the total falls, it was creating or defending it, and you now know which programmes to protect.

A channel that reports on its own performance will always look like the best channel you have.

What to do when the answer is uncomfortable

Say the test comes back and the channel everybody defends turns out to be harvesting. The instinct is a dramatic reallocation, and the instinct is wrong. Cutting a harvesting channel to zero usually costs you conversions you would have kept, because harvesting is still work and somebody has to do it.

Reclassify it instead. A harvesting channel is a cost of sale, not an acquisition programme, and it should be budgeted like one: an efficiency target rather than a growth target, sized against the demand that exists rather than the demand you want. This also dissolves something teams find mysterious, which is why pouring budget into a harvesting channel produces nothing. You cannot buy more of a demand nobody created, and that is the mechanism behind why paid search stops working when you scale the budget.

The other half of the fix sits on the page. If discovery channels are doing the persuading, a landing page arriving at the end of that journey does not need to sell from a standing start. It needs to be easy to act on, and it should ask for the smallest thing that moves the person forward, which changes the order of what you request and when: what a landing page should ask for covers that in detail.

It changes how you judge the people running the channels, too. Anyone reporting only their own platform’s numbers is grading their own homework, however honestly they do it. The fix is to agree the measure before the work starts rather than after the first report, which our parent company sets out in judging an agency without spec work.

Where we would start, and when this reverses

If you have a week, do the arithmetic that costs nothing. Add up every channel’s self-reported conversions for last quarter and compare the total against the number in your accounting system. The gap is the size of the double-counting. In most reporting stacks it is large enough to end the argument without anyone running an experiment at all.

If you have a quarter, run one holdout on brand search, then one on retargeting, in that order, and hold the line on the waiting period when somebody asks for an early read. They will ask in week two.

The advice reverses in one case, and it is a real one. If you are early enough that almost nobody has heard of you, there is no ambient demand to miscredit, and the measurable channels genuinely are creating most of what they report. Attribution errors are a problem of scale; they arrive with recognition. The signal to start testing is the moment branded search begins growing on its own, because that growth is the first evidence that something unmeasured has started doing the work.

Most of this is a reporting design problem before it is a media problem, which is why we treat strategy and analytics as the thing that decides what performance marketing is allowed to claim, and why a growth engine is planned as one system rather than a stack of separately reported channels. If a channel on your dashboard has been quietly winning for two years and nobody can explain the mechanism, tell us what it says and what your accounts say.