What You Are Actually Buying

For twenty years we taught clients to stop asking where their advertising would run. ChatGPT’s advertising business is the third time that question has changed, and this time it is much harder to check the answer.

The first time I put streaming video on a media plan, the client wanted to know which networks we were buying.

Then they asked which programs, and which dayparts.

We were not buying any of those things. We were buying impressions against an audience we had defined, and the system would decide where and when to find them.

That was a difficult thing to explain to someone who had spent an entire career believing that responsible media buying meant knowing exactly where the advertising would run. Television had coordinates. A network, a program, a daypart, and a rate card you could argue about on a Friday afternoon. What we were proposing had none of that.

We were buying the person, not the place.

Soon after, the change showed up in the budget. We recommended taking money out of broadcast, outdoor, and print, and putting it into search. The client was not happy, and I understood exactly why. We were cutting the media they could see in order to fund something they could not. But search reached people at the moment they were looking; it brought in customers we had never reached before, and it connected the investment to actual sales. We made the case. The money moved.

We were right. I have never had to apologize for that recommendation.

So I have spent twenty years teaching clients to stop asking where. Which is why it took me two readings to notice that the question has changed a third time, and that this time I am the one who wants to ask it.

Not an announcement, an arrival

ChatGPT advertising did not begin last week, whatever your feed told you.

It began as a United States test on February 9 of this year, for logged-in adults on the free tier and the low-cost Go tier. Through the year it widened. The United Kingdom, Mexico, Brazil, Japan, and South Korea by August 11. Then thirty-one European markets. On August 27, it reached India, with WPP and Omnicom named as the first agency partners, more than fifty brands live in the opening week, and a self-serve manager opening on September 4. Plus, Pro, Business, Enterprise, and Education carry no advertising at all.

That is not an announcement. That is an arrival. A test became a global media channel over seven months, while our industry was still deciding what to call it.

The Indian pricing is worth a moment on its own. The least you can spend to advertise inside ChatGPT is 725 rupees a day. The subscription that removes those same advertisements costs 399 rupees a month. So it costs roughly twice as much per day to be the advertisement as it costs per month to never see one. I have been doing this a long time, and I rarely get to watch a market price both of its sides that, honestly.

Prabhjeet Singh, who runs OpenAI in India, led with the guardrails rather than the opportunity. “Answers stay independent. Ads will be clearly labelled and not influence ChatGPT’s responses.” The corporate line is blunter: advertisers cannot pay to influence the answers ChatGPT gives.

I believe them.

I am not being arch. I think the claim is true as written; I expect it to stay true, and I think the people currently sharpening their knives for that fight are about to spend a year on the wrong war. Google said a version of the same sentence for twenty-five years, and it was also true. Paid results never rewrote the organic ten. That was never the thing that changed.

What changed is in the mechanics

The detail worth your afternoon is not in the press release. It is in the description of how a placement gets matched.

OpenAI lists conversation context, user intent, the advertiser’s landing page, the headline, the copy, and something it calls context hints. It then states, in its own words, that this is not exact match keyword buying.

Read that again as a buyer rather than as a reader.

A keyword is a thing a person typed. It belongs to them. For twenty-five years, when our industry bought intent, what we actually bought was a word some human being chose, and every auction was a fight among strangers over that word. It was crude, and it was public, and it had one great virtue: the unit of sale was evidence. Someone wanted something badly enough to type it.

A context hint is a description an advertiser writes of the situation it would like to turn up in. The model reads the conversation and decides whether this is that situation.

So look at where we have landed. For most of my career, the unit of sale was a place. Then it became a person. Now it is neither of those things, and there is no name for it yet, because what an advertiser is buying is the model’s reading of what the customer meant.

Which makes my old client’s question, the one I spent years training people out of asking, a rather better question than I gave it credit for. He wanted to know where the money went. He was not being sentimental about television. He was being an auditor.

The strongest case against me

Here is the honest version of the other side, and it is not weak.

Search did exactly this, at planetary scale, for a quarter of a century. It labeled its advertising. It kept it out of the organic results. The web got paid for, most people were not harmed, and an entire profession, mine very much included, built a good living inside the arrangement. On that record, conversational advertising is search with better manners, and a columnist wringing her hands about it is a woman who has forgotten how the last twenty-five years actually went.

I have some sympathy for that view. I have made money on both sides of it.

But one structural difference does not go away with better labeling.

In search, you could see the ten results the advertisement was sitting beside. You could not always be bothered, and most people were not, but you could look. The sponsored line was priced, by the reader, against the nine free things arranged around it. That comparison was the discipline. It is the reason the label meant anything at all.

An assistant returns one answer.

There is no visible set of alternatives beside it, because the alternatives were weighed and discarded upstream, inside a process the reader cannot see, and the advertiser cannot audit. Nothing has been corrupted here, and the answer really is independent. What has gone is the reader’s ability to check, and a label does not give it back, because the label tells you what the advertisement is and says nothing about what the answer was compared against.

That is not a scandal. It is a change in what a media buyer is responsible for, and I do not think our side of this business has understood that yet.

The part your dashboard will not show you

Here is the stake, and it will not surface in a quarterly review for three years.

Search kept the two rails apart. Organic performance and paid performance were separate disciplines, run by different people who frequently did not like each other. In an answer engine there is one rail. The record a model reads to decide whether to recommend you unpaid is the same record it reads to judge whether your paid placement belongs in that conversation.

Which means the thin, neglected, unowned public record of your brand is now setting your media price, quietly, in a channel that will never report it to you as a line item. Two companies will buy the same placement, in the same conversation, on the same day, and one of them will pay more, and the invoice will not explain why.

I wrote here in June that your brand has a narrator it never hired, and in July that the narrator has developed an opinion. This is the third thing, and it is the one with a number attached. The narrator now has a rate card, and everything you have published over the last five years is being negotiated on your behalf, whether or not anyone in your building is watching.

What we would do on Monday

Nothing exotic. Three things, in order.

Go and read your own record the way a model reads it, rather than the way your brand book describes it. Ask the major systems what your company does, who it is for, and who competes with it. Keep the answers in a document with a date on it. That document is now a media asset.

Second, before anyone buys a single conversational placement, write down what your brand is legible for. I do not mean the positioning statement. I mean the three or four situations in which a machine, with no reason to be kind to you, would treat you as the obvious answer. If no one in the room can write them, you are not ready to bid, and the auction will price that ignorance for you.

Third, put someone senior in charge of the record itself, and give them a budget. The discipline has two names, and the industry has not settled on either. Answer engine optimization, AEO, is the work of making a brand legible to the systems that answer questions rather than to the ones that rank pages. Generative engine optimization, GEO, is the same craft described from the other end: shaping what a generative system produces when it talks about you. At Optima IQ™, we treat both as media work rather than communications work, because they behave like media. They compound, they decay the moment you stop, and they show up in what you pay. Growth is Our Discipline™, and this particular discipline is unglamorous and completely learnable.

The grain ship

Cicero sets out a case in the third book of On Duties that I have not been able to put down since August 27.

A merchant sails from Alexandria to Rhodes during a famine with a hold full of grain. Prices are high. He alone knows that several more grain ships are close behind him and that the price is about to collapse. Is he obliged to tell the Rhodians?

Antipater of Tarsus said yes. Everything must be disclosed, so that the buyer is in no way ignorant of anything the seller knows. Diogenes of Babylon said no, or not quite: keeping quiet is not the same as lying, and a merchant is not required to argue against his own cargo.

Two thousand years later, our entire disclosure regime is built on Diogenes. Label the advertisement, tell no lies, and let the buyer look after himself. It has worked, and it works still, for exactly as long as the buyer can see the harbor.

What I keep turning over is the smaller question underneath, which is what becomes of the Rhodians when they can no longer count the ships in the water. Not because anyone lied to them. Because the harbor is now someone else’s to describe.

So: when a machine tells your customer that you are the answer, what did it read to get there, and who at your company is responsible for what it found?

By Ingrid Reyes, Founder and CEO of Optima IQ™.