Last Updated on July 7, 2026 by Taya Ziv
A few months ago, running an ad inside ChatGPT cost you two hundred thousand dollars before you wrote a single word. That was the floor. OpenAI confirmed it to Adweek, and for a while some reports put the real number closer to a quarter million. It was a velvet rope, and the bouncer was a wire transfer most founders will never make in the life of their company. So the whole thing felt like somebody else’s party. Big brands, big agencies, big budgets. Not you.
Then OpenAI took the rope down. In April the minimum dropped to fifty thousand. In May the self-serve Ads Manager went live and the minimum spend was removed entirely. The head of advertiser partnerships said the threshold was simply going away. Any advertiser in the US can now walk up, plug in a card, and buy attention from a product that reaches more than nine hundred million people a week. The party is open. Everyone’s invited.
And I want to tell you not to go in.
What actually happened
Let me be precise, because the numbers matter and the story moves fast.
ChatGPT switched ads on for the first time back on February 9. Sponsored results below the answer, labeled, not changing what the model says. That alone hit roughly a hundred million dollars in annualized ad revenue in about six weeks, which is a frankly ridiculous ramp for a brand-new channel. Then came the on-ramp changes. The two-hundred-thousand-dollar beta gave way to a fifty-thousand-dollar self-serve tier in April, reported first by Digiday, and by May the floor was gone completely. The Ads Manager is now a button, not a negotiation.
Here’s the number under the number, the one that tells you why OpenAI is doing this at all. Around twenty percent of everything people type into ChatGPT has commercial intent. One in five conversations is, in some form, a person trying to decide what to buy or who to use. That is the richest pool of purchase intent built since Google’s search box, and OpenAI just made it self-serve. Of course they dropped the minimum. They want the long tail. They want you.
Why this should make a founder pause, not pounce
The reflex is obvious. A new ad channel opens, it’s full of buying intent, the early auction is supposedly cheap, so you rush in before the prices go up. I get the instinct. I even wrote a version of it myself when ChatGPT first turned on pay-per-sale ads and it really did look like the AdWords moment all over again. That window was real. But here’s the honest update, and it’s the part nobody running an agency will tell you. The window that mattered was the one where almost nobody could get in. The moment the door opens to everyone, the cheap part is over.
Think about what an ad auction actually is. It’s a room where the highest bidder wins the slot, every time, forever. When the room was gated behind two hundred grand, a small advertiser who snuck in had a quiet auction with few rivals. Now that the room holds everyone with a credit card, you are bidding against companies with a hundred times your budget for the same one-in-five commercial query. You will not out-spend them. You never could. An open ad auction is the single most reliable machine ever built for converting “I have more money than you” into “I get the customer instead of you.” Walking in there at pre-seed is walking into the one fight where your scrappiness counts for nothing and their balance sheet counts for everything.
So no, the gift isn’t the ad. The gift is that the ad existing at all tells you exactly where the buyers now are. The trick is to reach them from a spot the auction can’t price you out of.
The real edge moved one line up the screen
Look at where a ChatGPT ad actually sits. It’s below the answer. The model says “here’s what I’d use for that,” and then underneath, smaller, labeled sponsored, sits the thing somebody paid to put there. The reader’s eyes have already landed on the recommendation before they ever reach the ad. The recommendation is the prize. The ad is the consolation slot you rent because you couldn’t be the recommendation.
So be the recommendation.
When a founder asks ChatGPT “what should I use to do X,” and the model names three tools, being one of those three names is worth more than every sponsored link beneath it combined. Nobody clicks “sponsored” with the same trust they give the answer the AI just handed them. And the wild part, the part that should make you put your coffee down, is that getting named in that answer is cheap right now. It’s gameable. It’s where search engine results were around 2010, before everyone figured out the rules and the easy wins dried up.
The clunky industry name for this is generative engine optimization, GEO for short, and there’s already a whole young category of startups built around the fact that most brands are simply invisible to ChatGPT. The early numbers are loud. At some companies, citations inside AI answers already influence close to a third of their sales-qualified leads. A YC-backed team out of Google and DeepMind, AthenaHQ, has a case study where a brand went from a two percent share of voice in ChatGPT answers to nearly thirteen percent in sixty days. Another customer roughly ten-x’d how often the model cited them. Profound, the firm leading this little category, raised over fifty-eight million dollars and counts Ramp and Figma and Zapier as customers, with Ramp reportedly lifting its AI visibility seven times in weeks. And the floor to start is almost nothing. There are tools that track whether ChatGPT recommends you starting at thirty-nine dollars a month. Thirty-nine dollars, against a fifty-thousand-dollar ad spend that was the entry price two months ago. That gap is the whole opportunity.
There’s a second floor being poured under this, too, and it’s worth watching. The agents aren’t just recommending anymore, they’re starting to buy. There’s an open Agentic Commerce Protocol now, Shopify quietly shipped Agentic Storefronts so a model can surface and sell your product inside a conversation, and Google is building its own commerce rail with Walmart and Shopify in the room. A YC company this winter, Sitefire, exists for the single purpose of helping you market to the agents instead of the humans. The shape of the next channel is forming, and it isn’t a billboard. It’s a recommendation that can complete the purchase without the customer ever leaving the chat.
The trap inside the good news
Now let me argue with myself, because if you take this as gospel you’ll get burned, and I’ve watched founders bet a company on a channel that turned out to be quicksand.
Cheap and gameable never lasts. The same sentence was true about Google in 2010 and Facebook in 2014 and every channel since, and the ending is always the same. The easy wins attract a crowd, the crowd attracts spammers, the platform tightens the rules, and the arbitrage closes. GEO will get gamed into the ground exactly like SEO did, and the model labs will start policing what gets recommended the way Google polices its results. There’s a darker version too, where OpenAI decides the recommendation slot is its own product to sell, and the free citation you earned this year becomes another auction next year. You’d be building on rented land owned by the same company selling the ads. That’s a real risk and I won’t wave it away.
And honestly, for some businesses the paid slot is fine. If you sell something with screaming commercial intent and a fat margin, and a sponsored line under a high-intent answer pays back on the first sale, then run it, cap it hard, and good luck. The auction isn’t evil. It’s just the wrong default for a founder whose only real edge is being early and clever instead of rich.
What to actually do about it
Three moves, and none of them is “wait and see how ChatGPT ads shake out.”
First, find out what the model already says about you, today, for free. Open ChatGPT and ask it the exact questions your customers ask when they’re looking for a tool like yours. See if you show up. See who does. That five-minute test tells you more about your real distribution in 2026 than your Google rank does, and almost nobody on your competitive set has bothered to run it.
Second, make yourself the easy answer. Put your real positioning, your specific use cases, and honest comparisons where a model can read them, and earn mentions on the third-party sources these engines actually trust, the same instinct behind getting yourself into the thin stream of AI search traffic that converts many times better than the ordinary click it replaced. You’re not writing for a human skimming a results page anymore. You’re writing to be quoted by a machine that’s about to vouch for you.
Third, get on a rail that lets an agent actually transact you, not just name you. If you sell anything a model could buy on a customer’s behalf, look at Shopify’s agentic storefront and the commerce protocols forming around it now, while being early still counts for something. The founders who were on Amazon’s marketplace in year one know what it’s worth to be standing on a platform before it gets crowded.
The line everyone’s about to stand in
Here’s what I actually believe. An ad is rented attention. You pay, the attention shows up, you stop paying, it vanishes, and the price only ever goes one direction once everyone can bid. A recommendation is borrowed trust, and right now it’s borrowed cheap. The whole industry is about to line up at the ad button OpenAI just opened, bidding each other’s prices up for the privilege of sitting underneath the answer. The smarter, lonelier move is to climb one line up the screen and become the answer, while that climb is still mostly empty.
The party’s open. Everyone’s going in the front door to buy a drink. The thing worth having is to be the name the host says when a guest asks who they should talk to. That costs almost nothing today. It won’t for long.
