Moving budget into the best performing channel reduced total sales. The reasoning behind the move was standard, the reporting supported it, and the outcome was the opposite of what the reporting predicted.
The reasoning that looked sound
Search consistently showed the strongest return of any channel in the account. Cost per acquisition was the lowest, and the relationship held across every reporting period examined.
Upper-funnel video and paid social showed weak direct returns. They generated impressions and little attributable revenue.
The obvious conclusion was to move money from the weak channels to the strong one. Almost every budget review reaches this conclusion at some point.
The move was made gradually, over a quarter, which is the responsible way to do it and did not help.
Search spend rose, search revenue rose slightly, and total revenue fell.
Search harvests demand rather than creating it
A search ad is shown to someone who has already formed an intention specific enough to type. The channel does not create that intention, it intercepts it.
This is why search returns look excellent. It is being paid for conversions that were substantially decided before the ad appeared.
Some portion of those conversions would have happened anyway, through an organic result or a direct visit. That portion is pure cost.
The rest depended on something earlier having created the intention, which is the part the channel does not do and cannot be credited for.
Increasing search budget therefore has a ceiling set by how many people are searching, which is set elsewhere.
The lag before the decline
Nothing happened for several weeks, which made the early evidence look encouraging. Existing demand was still in the system.
People who had seen the upper-funnel activity before the cut continued to search and convert, and the search campaigns captured them efficiently.
The pipeline of new intention was shrinking at the same time, invisibly, because nothing was refilling it.
The decline appeared roughly a quarter later, which is close to the length of the consideration cycle in that category.
That lag is what makes the mistake so common. The feedback arrives long after the decision, attached to a different reporting period.
Branded search volume was the tell
The clearest early warning was in query composition rather than in revenue. The proportion of searches that included the brand name began falling.
Branded queries indicate that someone arrived at the search already knowing who they wanted. That knowledge comes from prior exposure.
Total search volume held up for a while because generic queries continued, but generic queries convert less well and cost more to win.
Blended search performance therefore drifted downward while the campaign structure was unchanged, which is difficult to explain from inside the channel.
Watching branded query share is cheap and provides a leading indicator that revenue reporting cannot supply until it is late.
Why last-click made the move look correct
Under last-click attribution the final touch receives all credit. Search is very often the final touch, because searching is what people do immediately before buying.
Channels that appear early in the journey receive nothing, regardless of what they contributed.
This produces a systematic bias in every budget review. The measurement method rewards the channel closest to the transaction and penalises everything upstream.
Multi-touch models redistribute credit and do not fully solve it, because they still only see paths that reached a conversion.
The channels most likely to be cut are the ones whose contribution is hardest to observe, which is not the same as the ones contributing least.
The test that would have caught it
The question that mattered was incremental. What happens to total sales if this spend stops, not what does the platform credit it with.
That question is answerable by holding out a region, or a randomised set of users, and comparing outcomes against a matched control.
Geographic holdouts are the practical form for most advertisers. Switch a channel off in a set of markets, leave it on in comparable ones, and observe the difference.
It requires patience, because the effect takes as long to appear in the test as it did in the live account.
It is still far cheaper than discovering the answer by cutting the budget everywhere and waiting a quarter.
Rebuilding the split
Restoring the previous budget did not restore sales immediately. The demand pipeline had to be refilled, and refilling takes as long as draining did.
That asymmetry is worth planning for. The cost of the mistake includes the recovery period, not just the affected quarter.
We rebuilt the split with an explicit floor on upper-funnel spend, set as a proportion of total budget rather than as an absolute figure.
A floor expressed as a proportion survives budget changes. An absolute figure gets cut first whenever the total is reduced.
Search budget was capped separately at the level where incremental spend stopped producing incremental clicks, which is lower than most accounts assume.
What we hold back now
Every channel review now separates two questions that used to be asked as one. Is this channel efficient at what it does, and would sales fall if it stopped.
The first question is answerable from platform reporting. The second is only answerable by withholding spend and observing.
We run a rolling holdout so that at any time some portion of the market is unexposed to at least one channel, which gives a continuous read.
The cost of that holdout is real and it is small relative to the size of the decisions it informs.
It also settles arguments that otherwise recur every planning cycle, because the answer stops being a matter of which attribution model somebody prefers.