The Quarter Our Best-Performing Channel Turned Out To Be Measuring Itself

In Q2 2026 we increased paid search budget twice for a B2B software client, over 11 weeks, because it was comfortably our cheapest source of qualified leads. Cost per qualified lead was roughly 40 percent below every other channel we ran for them. Then a junior analyst on our team asked why branded search impressions had grown in months when we had not increased brand spend, and there was no good answer to that question. The channel we had been funding was not creating demand. It was capturing demand that other work created, and our attribution model was handing it the credit.

Nothing was broken. Google Ads was configured correctly, GA4 was firing, and the numbers in HubSpot reconciled with both. The measurement was working exactly as specified, which is how it produced a confident wrong answer for a quarter. Most of our work is founder-led marketing, where the demand-creating activity is a person talking in public, so this particular blind spot was expensive in a way we should have anticipated.

What we thought we were funding

The plan was ordinary. Branded and near-branded search was cheap, converted well, and every model we had said so. We moved budget into it in April and again in May. The first increase looked like it worked. The second one is where the pattern broke: spend rose by about half and qualified leads rose by almost nothing.

We did what most teams do with 2 weeks and a flat channel. We rebuilt match types, expanded negatives, tested 4 landing page variants, and moved from target CPA to target ROAS and back. None of it moved anything, which in retrospect was the clearest signal available, because a channel with a genuine optimisation problem responds to at least one of those levers.

The gap was in what the model could see, not in what it did

Our attribution was not naive last-touch. We ran a position-based model that spread credit across the observable path. The problem is the word observable.

For this client the demand-creating work was a founder posting on LinkedIn and X, two podcast appearances, and a trade-press article that people read without clicking. None of those produce a tracked touch. The model divided credit among the touches it could see, and paid search was reliably the last one before a conversion, so paid search kept winning a contest the other channels were not entered in.

The model was not wrong about what it measured. It was wrong about what existed, and because it was internally consistent, every check we ran on it passed. We had a reconciliation process that compared platform numbers to CRM numbers and it agreed every month, which felt like validation and was actually just two instruments sharing the same blind spot.

When we finally pulled branded search impressions against the founder's posting calendar, the correlation was not subtle. Podcast episodes and posts moved branded search volume 2 to 3 weeks later, and paid search harvested the result. Across 6 months there were 4 clear step changes in branded volume and every one of them followed something a human did in public.

We had been increasing budget on the harvester and holding budget flat on the thing that made the harvest exist.

We confirmed it the only way that settles this kind of argument, which is by turning something off. We paused branded search entirely in 2 of the client's smaller markets for 3 weeks and left it running everywhere else. Total qualified leads in the paused markets fell by about a fifth, not by the 40 percent the attribution model implied we should lose, and organic branded clicks absorbed most of the difference within days. That gap between the modelled loss and the observed loss is the size of the credit the channel had been taking for work it did not do. I would run that test earlier next time, because it took 3 weeks and answered a question 11 weeks of dashboard analysis could not.

What we changed

Three changes. Only the first one really mattered.

First, branded search stopped being an acquisition channel in our reporting and became a demand indicator. It still gets funded, because capturing your own branded demand is cheap and a competitor will capture it if you do not. But it no longer competes for budget against demand-creating channels, because that was never a fair comparison. It was one line item taking credit for another.

Second, we changed what the monthly review opens with. Branded search volume and direct traffic now come first, as a pair, before any channel performance. Those two are the closest observable proxy we have for whether the unobservable work is landing. When they go flat, the creation side is underfunded no matter how efficient last-touch says the bottom of the funnel is.

Third, no channel gets a second consecutive budget increase on attribution data alone. The second increase needs a holdout, a geo split, or at minimum a deliberate spend reduction where we watch what happens. That rule is genuinely annoying and it has cost us some upside. It has also caught 2 smaller versions of the same error since, both times within a month rather than a quarter.

What I would tell someone running this

The failure was not a bad model. It was trusting a model that had no way to represent the thing that mattered, and having no check that would ever have revealed the omission. We did not understand at rollout that our reconciliation process could only ever confirm agreement between two systems that were both blind in the same direction.

If you want one practical test, it is the one our analyst stumbled into. Find a metric your attribution model does not feed. Branded search volume works. So does direct traffic, or the share of inbound conversations where the person already knows what you do. Then ask what moved it. If your reporting cannot answer that, your channel numbers describe capture and tell you nothing about creation, and you will keep moving money toward the end of the funnel because that is the only part your instruments can see.

We now hold the creation side to cost per qualified outcome rather than to attributed conversions. It is a worse number in every respect except the one that counts, which is that it does not lie to us about causation.

The 11 weeks cost real money. The habit of asking what the dashboard is structurally unable to show has since saved considerably more.