Ad paths diverged

Breakfast table. The car. A hand held out with a screen attached to it because there’s a free app, or a social network everyone else is already on. Twenty years of it, and the refrain never changed, and nor did the eye-roll that met it. Repeat after me: if something is free, the price is me.

My children would recite it after me, until they just started it with the ask itself: “Dad, I know that if something is free the price is me, but…”

This month OpenAI wrote it into a price list, and I want to use it to say something about advertising, not about free apps.

For two years I have written about this industry splitting in two: integrated scale at one end, specialists at the other, and a middle that is dying not because companies are moving into it but because its edges keep moving out (Crossing the Deathly Middle, two weeks ago).

The other split is one I have not written about, and it is the one in the product. Advertising itself is bifurcating, and ChatGPT has just shown us which side it has chosen.

At one end sits advertising that is genuinely useful: the right message to the right person in the right context, because that person has allowed themselves to be known, or has told you where they are on a journey. The purest form of it is Google Search. I type a question. I know that among the answers will be some that people have paid to put in front of me. I asked the question. I asked for the answers. I am not offended that one of them cost someone money to be there, because it is still an answer to the thing I asked. That contract has served the trade for twenty years, and to an ad man it is the cleanest thing we ever built: advertising as a reply to a question that was actually asked.

At the other end is the advertising that interrupts, distracts and sprays. It gets between people and what they were trying to do. It lives wherever people cannot afford, or do not choose, to pay to make it stop.

A hard truth for anyone who sells advertising for a living is that on screens, the second kind is now the only kind most people meet, because the first kind is increasingly behind a paywall of a different sort. Streaming has a tier where the ads stop. News has one. Music has one. The ads flow to whoever won’t pay.

Last week OpenAI said ChatGPT Ads are now live in 89 countries, that more than 100,000 advertisers have run a campaign since US testing began in February, and that the business passed a $1 billion annualised run-rate at the end of August (a big number, until you remember that Google’s advertising business did $82 billion in the last three months of 2025 alone).

And it restated who sees them: users on the Free tier and the $8 Go tier. Not Plus, Pro, Business or Enterprise. Not anyone under eighteen. Not near health, mental health or politics. Pay twenty dollars a month and you will never see one.

Google never had that tier. The chief executive and the student typed the same words into the same box and saw the same sponsored answer, and that sameness was the democratic part. Here, the moment you pay, you leave the audience.

Before an advertiser sets a bid or writes a word, the people it will reach have been sorted by one signal it did not choose: who would not, or could not, pay to avoid it. The price is me, printed on a price list.

In fairness, if any medium was born to sit at the useful end, this was it. A ChatGPT conversation is a question, in the user’s own words, with the context attached: not a keyword, the whole problem. OpenAI’s January principles are a defensible starting point, too: ads do not influence answers, conversations stay private, no data is sold to advertisers, nothing near the vulnerable or the sensitive. The auction is relevance-weighted, so a well-matched ad with a small bid can beat a lazy one with a big bid, which rewards craft. The product is eight months old, and search did not have match types or a search-terms report in 2003 either. When OpenAI published partner results this month, WeightWatchers reported a 15.3% lower cost per acquisition. Their number, but a number. And a hundred thousand advertisers say people are finding something.

All of which makes the product it actually built harder to understand, and I spent last week trying to understand it, as an advertiser, with the Ads Manager and the help pages open in front of me. What is it that I am buying?

It is not search. There are no keywords, no match types, no buying of the question. You write “context hints” in plain language describing the conversations you would like to be near, and OpenAI’s own documentation says hints guide matching but “do not guarantee delivery.” It is not what Perplexity was offering, which wove sponsors into the answer itself and then, in February, left advertising altogether, citing user trust. What it is becomes clear when you ask the three things a media buyer could ask Google in 2004.

Where did it run? Device, platform, country. Nothing finer than country. That is the whole breakdown.

What did they ask, and why did I appear there? You are not told. Not the prompt, not the answer, not an inferred topic, not which of your hints matched. The reporting is impressions, clicks, spend and conversions, and OpenAI’s stated basis is that advertisers receive aggregated, non-identifying performance data only. You will never know which question your brand was offered as the answer to.

What did it say? Until this week, your words. From Friday, possibly not. Last week OpenAI emailed self-serve advertisers that “creative enhancements” will be switched on for every new and existing campaign at the end of this week unless they opt out. Under one setting OpenAI may rewrite your headline and description to fit the conversation, translate the ad, resize the image and send the click to a different page on your site than the one you chose. Variations may be prepared in advance or generated in real time. The help page is candid: “You do not need to review or approve each enhanced version before it runs.” Nor can you. There is no preview of every variation, the reporting does not separate enhanced versions from the original, and OpenAI’s own caveat is that a change in your metrics does not by itself tell you what the rewriting did.

Much of the pricing history, the $60 launch CPM and the $200,000 minimum that fell to $50,000 and then to nothing, comes through the trade press rather than from OpenAI, and the one independent study on relevance I rely on next I have read only through analysts’ accounts.

That study, by SE Ranking as EMARKETER reported it in August, found 14.4% of ChatGPT ads had no topical connection to the question asked, and more than half of those served in news and politics conversations. Which is the point. The ad sits beneath the answer, looks like a display unit, and may or may not be relevant to the question above it. You did not buy the question. You cannot see the question. From Friday you may not even have written the words. One practitioner who has tried to run it seriously put it better than I can: relevance you cannot verify is a position, not a measurement.

So the one medium that holds the question in writing, in the user’s own words, with the context attached, has chosen to sell the space under the answer, to the people who did not pay to make it go away. That is the spray end of the spectrum, dressed in the clothes of the useful end. The holding companies will sell it all the same; Omnicom, WPP, dentsu and Publicis are all resellers, and on Thursday’s earnings call Omnicom named AI-driven discovery as a growth area. It will be. The question is whose.

What would I do, as the advertiser I was last week? Opt out of creative enhancements before Friday, at account level, and opt back in deliberately if a test earns it. Treat the buy as reach, not intent, until there is a report that tells me what intent I bought. Ask the agency what it can see that I can’t, because if the answer is nothing, the reseller margin is for the logo. And use the one lever that still points at the conversation, the hundred negative phrases, because it is the only piece of long-tail discipline the product allows.

But the thing I want to say is not advice. It is that this makes me sad, as an ad man, and it is this:

Not once, finishing a prompt in any of these tools, have I hoped there would be an ad under the answer.

Nobody has. And the reason nobody has is that we put a price on avoiding it and, in doing so, taught people that advertising is a thing you pay to avoid, when the best of it, for twenty years, was the thing you asked for. ChatGPT had the chance to be that. It had the question. It chose the footer, and it chose the audience that couldn’t leave.

My children were right to roll their eyes. The line was obvious. It was also true, and now it has a tier sheet.

The one medium that holds the question in writing has chosen to sell the space beneath the answer.

ChatGPT knows the question, in the user’s own words. If it sold you the question instead of the footer, would you buy it, and which kind of advertising would that be?

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry’s table: senior client, global agency leader, technology founder. It is read twice weekly by senior marketing and agency leaders, and it is the kind of advertising you asked for. Every Monday and Thursday: brandflow-marketing.com.