OpenAI is rolling ChatGPT ads into 31 new markets, on top of the five where they already run, and its business CMO says daily ad revenue has grown 25% since the start of August. That is a land grab timed against a Q4 IPO at a reported $852 billion target. I have watched this exact sequence four times in 22 years of buying media, and the part worth acting on is not the product. It is the price.
Every channel I have bought into early was cheap for the same reason: the inventory existed before the demand did. Search in the mid-2000s, Facebook before the pixel matured, YouTube before brand safety tooling, Amazon before the agencies staffed for it. In each case the spread between what the inventory cost and what it was worth stayed open for somewhere between eighteen months and three years, and then closed permanently. Nobody rang a bell at the top of any of them.
What scarcity looks like once the spread closes
The other end of that curve is on display in sports right now. Disney and ESPN announced a Super Bowl sellout roughly a month earlier than NBCUniversal managed the year before, with 58 brands across 34 categories on the board. A 30-second spot ran about $8 million last year.
That same 30 seconds cost $37,500 in 1967.
Nothing about the audience improved by a factor of two hundred. What changed is that the inventory became provably scarce and the buyers became certain about its value at the same time. Certainty is what advertisers actually pay for, and it always arrives late.
Why a citation behaves like an impression
The reason I take the ChatGPT expansion seriously has nothing to do with the ad units themselves, which are early and unremarkable. It is that the surrounding behaviour has already changed. Meltwater analysed 9.5 million AI citations across six models and found LinkedIn’s citation share rising 26% in a four-week window, with 48% of cited content published in the previous three months.
Read those alongside the 25% ad-revenue growth and a shape appears. Buyers are moving their question-asking into assistants faster than budgets are following them there. Attention has relocated; spend has not. That gap is the entire opportunity, and it is the same gap that made every earlier channel cheap.
The unit is changing too. For twenty years the atomic unit of paid attention was the impression: a placement, counted, priced against a thousand of them. In an assistant, the equivalent event is a citation. Something gets named in the answer a buyer reads. It is countable, it is attributable, and unlike an impression it carries an implicit endorsement from the model doing the summarising. I do not think the industry has priced that endorsement at all yet.
What I am actually doing about it
In my own accounts the move has been deliberately small, because a land grab is not a reason to abandon channels that work. Concretely: a test budget that would be a rounding error against the main search spend, aimed at the markets where the client already has demand rather than the new ones, and measured on assisted conversions rather than last click, because a channel this early will never win a last-click comparison against branded search.
On one e-commerce account I run, the instructive part was not the return. It was that the incremental cost of finding out was smaller than a fortnight of the same client’s Shopping budget. When the cost of information is that low relative to the cost of being late, the argument for waiting has to be very strong, and usually it is just discomfort wearing a spreadsheet.
The part I cannot support
I called this underpriced in the title and I should be precise about what that claim rests on, because it does not rest on a rate card. I have not seen OpenAI’s CPMs, no meaningful benchmark set exists across 31 markets that launched this month, and any number quoted at me privately would be a sample of one advertiser in one vertical. What I am reasoning from is the shape: relocated attention, revenue growing fast off a small base, a company with a strong incentive to show advertiser traction before an IPO, and no established buying discipline on the other side of the table.
That shape has been right before. It has also been wrong. The obvious way this ends badly is that assistant advertising turns out to be structurally different from feed advertising, because a user in a task-completion mindset resents interruption more than a user who is browsing. If that is true, the inventory is not underpriced, it is correctly priced for something that converts poorly, and the early buyers subsidise everyone else’s learning.
A second failure mode is regulatory. Disclosure norms for AI-adjacent advertising are unsettled, and a channel that has to retrofit consent or labelling can lose its economics overnight.
The test that would change my mind
What would move me is boring and specific: whether assisted conversions from assistant-sourced sessions hold up at three months rather than spiking and decaying. If they decay, the citation is a novelty impression and I have overread it. If they hold, the spread closes, the price goes up, and the people who bought early will describe themselves as visionary when in fact they were just early.
My position is that being early costs less than being certain, and certainty is exactly what pushed a 30-second spot from $37,500 to $8 million. The only thing I am confident about is that nobody will announce when the window shuts.