The clicks were cheapfrom the first day.Nothing else was simple.
A subscription app in a crowded consumer category started buying ads inside an AI assistant while the format was new enough that nobody had a benchmark for it. This is an anonymous read: the client is not named, and every figure is a change rather than an amount. It is also not a win. Four months in, the channel still moved every time it was touched.
- Over a third
- cheaper per click than search on the same account
- Before 6am
- the hour the whole daily budget ran out
- Close to half
- off the cost per subscriber, first reading to last
- Still moving
- the channel had not settled when these readings stopped
Cheap traffic that would not finish the job.
No benchmark, in either direction
The team had no benchmark to aim at, because the format was too new for anyone to have one. That cuts both ways. There was no inflated expectation to manage either, so the first weeks went on establishing what normal was rather than on defending a number.
The budget was gone before the working day
Spend cleared in the small hours and stopped, so the part of the day when this audience is awake had no cover at all. The platform was not pacing the budget across the day, and nothing in the interface said so.
The question carried intent. The click did not.
Someone who has just typed a paragraph describing what they want is about as warm as traffic gets. By the time that becomes a click on an ad and a landing page that opens the conversation again, most of what they told the assistant is gone. Cost per click stayed low. The funnel behind it did not convert at the rate the click price implied. The bottom of that funnel was not the problem: once somebody started onboarding, they reached the paywall at close to the same rate as traffic from search. With the click materially cheaper and the last step unchanged, the whole of the difference has to sit in the step in between, which is the one nobody instruments.
Catch the intent before it evaporates.
Traffic sent where the intent could be captured
Into the funnel directly rather than through a landing page, with one open question at the top. If someone has already described their problem to an assistant, the cheapest way to find out what they said is to ask them once, immediately, rather than to infer it from a click.
The day given cover
Pacing fixed so the budget lasted past the morning, and daily spend raised in steps rather than in one move. Each step was allowed to run long enough to be read before the next one.
The structure left alone while volume was thin
Bids raised where an ad group flattened, but the multi-group structure kept until there were enough conversions to justify changing it. Restructuring an account on low volume resets what the system has learned and teaches you nothing in return.
Copy rewritten where click-through lagged
The gap between what somebody asked and what the ad offered is a copy problem before it is a targeting problem. Where the two did not match, the copy changed first, and the targeting was left to prove itself afterwards.
Two readings, three weeks apart. They are not a trend.
Cost per subscriber is shown as a proportion rather than in currency, and the figures on this page are bands rather than point readings: the first reading is the baseline, the second is what was left of it. You see the size of the change without seeing the amount, and without a precision that would help anybody work out whose account this was. Both readings come from the account team looking at the panels on a dated call.
Two points are a line, not a trend, and this line is the flattering version of a month that was not flat. A reading between these two was lower than either, and a written summary from the same month put the cost higher than the week before it. By mid-September delivery was unstable again. So the honest summary is that the cost per subscriber moved by roughly half across three weeks and had not settled, which is a different claim from the one a chart with two bars invites you to make. There is also no incrementality test behind any of this. One was scoped for this channel and never ran, so nothing here separates subscribers the channel created from subscribers who would have arrived anyway.
Buy the intent, then do not throw it away.
A question carries intent and a click does not, and the distance between the two is where the money goes. If you buy this traffic and land it on a page that starts the conversation from the beginning, you have paid for somebody who already explained themselves and then asked them to do it again. Ask once, immediately, in the first thing they see.
Pick the AI surface you can measure, not the one with the better story. One of them passes your tracking parameters through and one of them does not, and that difference decides whether you will be able to say anything about the channel in six months. It is a duller reason than the usual ones and it is the one that holds.
Judge it over a window that matches how long your customers take to decide. On a long cycle a short measurement window will understate the channel, and a short window is what most accounts use, because it is what the platform shows by default.
This work was delivered by MTA Group, which Zero Fluff Digital is part of, by the same specialists who would work on your account. The figures on this page are readings taken by that team from the advertising and analytics panels during the engagement, recorded on dated calls. They are not an export from either platform. The client is not named and the figures are given as changes rather than amounts, at their scale rather than in their currency. There is no third-party study to link to here, so treat this as weaker evidence than the cases that carry one, and ask us for the workings on a call. The reviews behind our work are public and verified on Clutch, where clients rate the work rather than the agency describing itself.
Measured between May and September 2026.
Thinking about buying attention inside an assistant?
30 minutes, no deck. You find out whether this channel can reach your buyers at all, what it would cost you to learn that, and how you would know afterwards whether it worked.
Book a callNext case studyNegative keywords scored by a decision model: they stop 0.5% →