The channels nobody clicks

Every attribution report has a channel near the bottom that everybody privately suspects is doing more than the chart says. It is usually video or audio. It has spend against it, a small number of conversions, and a cost per acquisition that would be indefensible if anybody believed it.

Two explanations are available. One is that the channel is genuinely bad. The other is that the channel produces something the report does not count. Both happen, and telling them apart is the entire job — because the cost of getting it wrong runs in both directions.

A model that only sees clicks will always rank a channel that does not produce clicks last. That is a property of the model, not a finding about the channel.

What a video or audio ad actually does

Video and audio advertising mostly creates demand rather than capturing it, and demand shows up later, somewhere else, under another channel's name.

The sequence is unremarkable when written out. Somebody watches a pre-roll, or hears a spot between tracks, and does nothing at all — because they are watching something else, or driving. Days later they have a problem the advertiser solves. They search the brand name, click the paid search result at the top, and enquire.

Last-click reporting records that sale as branded search. It is not wrong: the click really did come from a branded search ad. It is simply answering the question "what did they click last" and being read as though it answered "what caused this". Those are different questions and only the first is observable.

The compounding version of this is worse. Because branded search looks efficient — cheap clicks, high conversion — budget moves toward it. Branded search cannot create the demand it harvests, so total demand stops growing, while the report continues to show excellent efficiency right up until growth flattens.

Nobody argues with a channel that is quietly losing money once the loss has a figure beside it. The reverse is also true: nobody defends a channel whose contribution has never had a figure beside it.

The measurable part, and it is not nothing

Every video or audio flight can be given at least one direct response route, and revenue arriving through that route can be attributed exactly as rigorously as search.

The instinct is to treat these channels as wholly unmeasurable and reach for modelling immediately. That skips a step. Most flights can carry a direct signal at almost no cost, and a small verified number is worth more than a large modelled one.

That last row deserves more respect than it usually gets. A single optional question — how did you hear about us — placed on a form that people already complete, produces a large sample of self-reported attribution at effectively zero cost. It is noisy. It is also the only signal that captures somebody who saw a video, told a colleague, and had the colleague enquire.

None of these routes captures everybody. That is the point to make explicit rather than to hide: what they produce is a floor. If a flight produced forty attributable sales through a vanity URL and a code, the honest sentence is that it produced at least forty, not that it produced exactly forty.

A verified floor beats a modelled estimate, provided you say the word floor out loud. The failure is not undercounting; it is undercounting silently.

Reading the indirect evidence without fooling yourself

Changes in branded search, direct arrivals and total lead volume across a flight are real evidence about a channel, and they are evidence rather than attribution — the distinction matters and should survive into the report.

The pattern to look for is boring and reliable. During and shortly after a flight, branded search volume rises, direct arrivals rise, and overall lead volume rises by more than the directly attributable amount. When the flight stops, those figures decay over a few weeks.

This is genuinely informative. It is not proof, because other things happen in the world during a flight — a competitor's outage, a seasonal peak, a press mention — and none of that is controlled for. Anyone presenting it as proof is overclaiming, and will eventually be caught overclaiming, which costs more than the modest claim would have earned.

Three ways this analysis goes wrong

1. The window is chosen after the fact. If you pick the flight window once you have seen the data, you will find a lift, because there is always some window that contains one. Write the window down before the flight runs.
2. Seasonality is ignored. Comparing a December flight against November proves very little in most businesses. Compare against the same period a year earlier as well as the period before.
3. The lift is converted into revenue and added to the attributed total. This is the one that destroys reports. The moment a modelled figure joins the same column as verified revenue, the whole column becomes an estimate and nobody outside marketing will use it again.

Keep them in separate columns and the report stays usable by finance. Blend them and you have produced a document that is persuasive to people who already agree and worthless to everybody else.

What a defensible video report looks like

Three numbers and one stated limitation: matched revenue as a floor, the direct-route detail behind it, the movement in branded and direct volume across the flight, and an explicit note that branded search revenue has not been reassigned.

In practice the document is short. Here is the shape, and the shape is more important than the figures in it:

• Matched closed revenue from the flight's own routes, with the match rule stated — this is the number that reconciles to the ledger.
• The routes themselves, so a reader can see how thin or thick the direct signal was and judge how far below the truth the floor probably sits.
• Branded search and direct lead volume for the flight window, the preceding window and the same window last year, without any conversion of that movement into money.
• One sentence: revenue that arrived through branded search has been credited to branded search, and some of it was probably caused by this flight.

That last sentence is doing the most work in the document. It tells the reader exactly where the uncertainty lives, which means every other number in the report can be taken at face value. Reports that do not contain a sentence like it invite the reader to distrust all of it, because they can tell something is being smoothed over and cannot tell what.

CloseRev credits a sale only where a closed-sales row matches a lead or call on a normalised phone number or email. Everything else stays in Direct / Unknown rather than being spread across channels — which is exactly the bucket a video flight's indirect effect should be visible in, and visibly not claimed.

The test that actually settles it

A geographic holdout — running the flight in some regions and deliberately not in others — is the only method available to an ordinary marketing team that produces a causal answer rather than a correlation.

Everything described so far is either direct attribution, which undercounts, or a before-and-after comparison, which cannot separate the flight from everything else happening at the same time. A geographic split fixes the second problem by giving you a version of the same weeks with no advertisement in it.

The mechanics are within reach of any business operating in more than one area. Split your regions into two comparable groups, run the flight in one, and compare closed revenue per head across both for the flight window and a decay period afterwards. Seasonality, competitor activity and market conditions apply to both groups, which is precisely why the comparison means something.

Where geographic tests go wrong

• The groups are not comparable. Splitting by a business's largest city against everywhere else compares a market with different competition, different pricing and different demand. Match on size and historic performance, not on convenience.
• The control is contaminated. National media, organic social and word of mouth all cross regional boundaries, so a control region is rarely perfectly clean. This shrinks the measured effect, meaning a positive result is trustworthy and a null result is weaker evidence than it looks.
• The test is stopped early. A flight measured for less than one full sales cycle plus a decay period will show a smaller effect than the truth, and stopping when the numbers look good is how a real effect gets reported at twice its size.

Run one properly and you have something no platform report can give you: a number that survives somebody asking how you know. Run none, and the video budget is defended each year on conviction, which works right up until the year it does not.

Frequency, and the part of the budget doing nothing

Most of the waste in a video or audio budget is not the channel being wrong; it is a minority of the audience being shown the advertisement far more often than persuasion requires.

Reach and frequency distribute unevenly. A flight reporting an average frequency of four will typically have shown a small slice of the audience twenty impressions or more, and a large slice one or two. The average conceals both ends, and the expensive end is the small slice.

This matters for measurement because it changes what a poor result means. A flight that produced little may have been badly targeted, or may have spent most of its budget re-reaching people who had already decided. Those call for opposite responses — better creative or a different audience in the first case, a frequency cap in the second — and the average frequency figure cannot tell them apart.

Ask instead for the distribution: what share of unique viewers saw the advertisement once, two to four times, five to nine, ten or more, and what share of impressions each group consumed. If the top band is consuming a large share of the impressions, the first optimisation is a cap, and it is usually the cheapest performance improvement available on the whole plan.

None of this is attribution, and it should not appear in the revenue report. It belongs in the flight review, where the question is not what the channel produced but whether the money inside it was arranged sensibly. Keeping those two documents separate is what stops a revenue report from turning into a defence of a media plan.

Average frequency hides the waste. Ask for the distribution, and expect to find a small share of the audience consuming a large share of the budget.

The decision this is all for

The purpose of measuring a demand-creating channel is not to give it a fair share of credit; it is to avoid cutting something that works because it does not photograph well.

Consider the two errors. Cutting a video budget that was actually working produces a slow decline in total demand that is invisible for a quarter and then obvious for a year, and by the time it is obvious the causal link is impossible to prove to whoever approved the cut. Keeping a video budget that was not working wastes money at a visible, boundable rate.

Those errors are not symmetrical, and the asymmetry argues for measuring rather than for guessing in either direction. The middle path — attribute what you can verify, observe what you cannot, and never mix the two — costs one honest paragraph and protects against both.

It also changes the conversation with a platform. Rather than accepting a view-through window and reporting the output as revenue, you ask what identifier it can attach to a person who responds, and you build the flight around a route you can actually follow. Channels that can answer that question are measurable. Channels that cannot are still worth buying — you simply have to say so plainly, and hold them to the evidence you can gather rather than the evidence you would like.

Attribute what you can verify, observe what you cannot, and never put the two in the same column. That is the whole method, and it is what keeps the rest of the report believable.

Questions people actually ask

Why does YouTube look terrible in my attribution report?

Usually because the report counts clicks and YouTube's job is not to produce clicks. A video that persuades somebody to search your name a week later has done real work that a last-click model records as branded search. The channel is not underperforming; the model is measuring the wrong event.

How do I measure a channel that produces no click at all?

Two ways, used together. Capture whatever direct response it does produce — a dedicated landing path, a vanity URL, a tracking number, a promotional code — and attribute that normally. Then look at the lift in branded search, direct arrivals and overall lead volume across the flight, and treat that as evidence about the channel rather than as attributed revenue.

Is a promotional code or vanity URL good enough for CTV and audio?

It is the best available direct signal and it systematically undercounts. Only a fraction of persuaded buyers use the code, so the measured number is a floor rather than an estimate. Report it as a floor and say so — a floor that is honestly labelled is far more useful than a modelled figure nobody can check.

Should video and audio be judged on the same cost per acquisition as search?

Not on the same number, because they are doing different jobs at different points. Search harvests existing demand; video and audio create it. Judging a demand-creating channel by a harvesting channel's metric guarantees it loses, which is how businesses end up bidding only on their own brand name and wondering why growth stopped.

What is the minimum honest report for a video or audio flight?

Matched closed revenue from whatever direct route the flight offered, stated as a floor; the change in branded search and direct lead volume across the flight window; and an explicit statement that revenue arriving through branded search has not been credited to the flight. Three numbers and one caveat, all checkable.

Found this useful?

• Share on X
• Share on LinkedIn
• Share on Reddit
• Share on Email