Picture the Q4 planning meeting. Slide nine has one word on it, in a very large font: agentic. Somebody asks what the ChatGPT ads test actually returned. There’s a pause long enough to hear the air conditioning. Then the answer arrives, the one every marketer has given at least once: it’s early, but the signals are encouraging.
The signals are not encouraging. A working paper by Maximilian Kaiser and Christian Schulze, currently under review and covered by Digiday this week, looked at ecommerce sites and found that paid traffic from ChatGPT generated $0.89 in revenue per session for U.S. advertisers. Paid Google Search traffic generated $2.45. By our math, that’s about 36 cents on the dollar. Conversion tells the same story: 0.37% from ChatGPT against 1.17% from Google.
Meanwhile, the line on the other chart is going straight up. Adthena counted 1,082 brands running ChatGPT ads in June. By August it was 6,333. More than six thousand brands, in a single quarter, piling into a channel that, on the best data anyone has published, converts at roughly a third of the rate of the boring one they already own.
That isn’t a media strategy. That’s FOMO with a purchase order.
To be fair to OpenAI, the business is real, and it’s fast. Ads started testing in February. By the end of August the company said its ad business had passed a $1 billion annualized run rate about 200 days in, with a 2026 target of roughly $2.5 billion, according to SiliconANGLE’s coverage of OpenAI’s own post. Ads are rolling into more than 40 new countries. The product has around a billion weekly users. Any planner who ignores that entirely is being precious.
But revenue for OpenAI and revenue for you are two very different numbers, and the industry keeps confusing them. The first proves that people are buying. Only the second proves that it works.
There’s a quieter tell in the spend data, too. Sensor Tower, tracking U.S. ChatGPT ad spend since April, found shopping’s share falling from 38% to 20% by September, while financial services climbed from 2% to 13%. Shares aren’t totals, so nobody should read that as retailers fleeing. But the category the format supposedly suits best, people asking a machine what to buy, is no longer the one growing fastest inside it. When the obvious use case cools off first, pay attention.
Measurement, right now, is a swamp. Farhad Divecha, group CEO of Accuracast, told Digiday “We continue to see discrepancies there,” and suggested some of the underperformance may come down to broken attribution. Fine. That cuts both ways. If you can’t measure it cleanly, you also can’t claim it’s working, and you definitely can’t justify moving brand money into it on the strength of a hunch and a press release.
Then there’s the creative. On October 30, OpenAI began switching on what it calls creative enhancements by default for self-serve campaigns, new and existing, according to Search Engine Roundtable. The system can write new headlines and descriptions, resize images, swap destination URLs and translate copy, drawing on your landing page plus the context of the user’s conversation, sometimes generating versions in real time. OpenAI’s email to advertisers was refreshingly candid: “these ads can run automatically without requiring you to review or approve each version.”
You can opt out. But read that sentence again and ask who is writing your brand’s voice now.
For a brand that has spent years building a tone people recognize, a machine improvising your copy inside a stranger’s late-night chat about their landlord or their breakup is not an enhancement. It’s outsourcing your personality to the one party in the room with no stake in it. At least one paid-search veteran who tested the format early has already paused, citing lead quality. That’s a canary. Nobody is listening to the canary.
Now the counterargument, and it’s a decent one. Every new channel looks worse than the incumbent at first. Early paid search was messy. Early Facebook ads were a punchline until they weren’t. Early movers get cheaper attention and learn the platform before the auction gets crowded. If you wait for clean proof, you’ll arrive just in time to pay full price, and you’ll be learning the basics while your competitors are optimizing.
We agree with every word of that, and it’s exactly why the answer is a test budget, not a migration. Learning is a legitimate objective. Put a real number against it, small enough to lose and big enough to read, with a hypothesis written down before launch and a kill date on the calendar. Turn off the auto-rewrites so you’re testing the channel and not a robot’s idea of your voice. That’s how grown-ups explore a new platform.
What isn’t legitimate is the thing we keep seeing: brands pulling spend from the work that makes people want them, to chase a channel whose main proof point is that other brands are chasing it too. That’s not exploration. That’s bullshit with a dashboard.
Here’s the part that should make everyone a little uncomfortable. Think about what an AI answer engine actually does. Someone asks it for the best running shoe for flat feet, or a decent brunch spot in Echo Park that takes reservations, and it synthesizes an answer. The brands that show up, paid or not, are the ones it already associates with the question. The ones people already talk about, review, post about and search for by name.
In a world where a chatbot is the shopkeeper, being famous is the shelf placement.
That makes brand-building more valuable in an AI-mediated market, not less. Les Binet and Peter Field’s long-running work, as summarized by WARC, puts the optimal split at roughly 60% long-term brand building and 40% short-term activation, with the ratio varying by category. The ChatGPT gold rush is a lot of marketers sprinting toward the activation end of that spectrum, inside a platform whose answers lean on whoever the culture already knows.
Remember what OpenAI promised users when it announced ads in January: that their conversations are “never sold to advertisers,” as Fortune reported. Good for users. It’s also a reminder of what you’re actually buying. You’re renting a slot beside the answer. The answer itself still belongs to whoever owns the reputation.
This is why we build social work to be searched for, not just scrolled past. A post that gets saved, sent to a group chat and screenshotted creates the one thing no ad unit can buy: someone typing your name into a box of their own free will. That’s the signal every search engine, old or new, is built to reward. It’s slower to show up on a dashboard. It compounds anyway, and it doesn’t stop working the day the test budget runs out.
So test ChatGPT ads. Seriously. Learn the auction, learn the formats, write down what you find and share it with your team. Just don’t confuse a fast-growing advertiser count with evidence, and don’t let a 36-cent click talk you out of a dollar’s worth of brand.
Six thousand brands are in. Most of them will learn something expensive. The ones still standing in two years will be the ones people asked for by name, before the chatbot ever had to guess.
