OpenAI sells ChatGPT ads to over 100,000 customers as Anthropic keeps Claude ad-free
Wired's Model Behavior newsletter takes stock of advertising inside AI chatbots. Anthropic is holding to the pledge from its Super Bowl commercial this year to keep Claude ad-free, and its recent IPO filing did not mention selling ads at all. The author notes that this is an easy stance with about $112 billion in investor cash. The rest of the industry may not have that luxury: the piece argues that selling ads may be necessary to defray the high costs of generating AI responses.
OpenAI is all in. It has had over 100,000 customers advertising in ChatGPT since ads debuted there in February. Last week Google expanded ads in its search chatbot AI Mode to 10 more countries beyond the US and declared that users "find them highly relevant." Advertisers, by contrast, are far from satisfied with the experience, though they expect it to improve. The author adds that ads matter most for niche chatbot platforms, which cannot lean on the venture capital of the top AI labs or the revenues of tech giants.
To see what is working, the author spoke with Koah Labs, a San Francisco ad tech startup that markets itself as bringing ads to over 200 apps, among them DeepAI (tools for creative queries) and Liner (citations in AI answers). Its brands include telehealth company Hims and flower delivery service UrbanStems. Koah has about a tenth of the advertisers OpenAI has, but CEO Nic Baird says he has a good sense of what works.
Machine-written copy: Baird says ads composed partly by AI rather than fully by a human generate double the clicks. The source says it is hard to know why, but AI-written copy likely feels more natural beside an AI model's output. Because Koah's ads are grounded in product details and imagery supplied by the advertiser, users do not seem to be meeting AI-hallucinated, erroneous information. Advertisers such as drugmakers, who face government rules on what they must or cannot say, default to more human control and get weaker results, according to Baird. The author warns that dull or cluttered ads could drive users away, and that the balance between engaging and risk-averse may be harder to strike with no human in the loop.
Creative targeting: a Google search for "help with stress" is likely to bring up ads for therapists, while Baird says spa ads have done well in Koah's network after stress-related prompts. The difference, he says, is that Koah can use "not only the prompt, or the Google search, but also the response from the AI platform, which brings context from the rest of the conversation thus far." Chatbot prompts also tend to be at least double the length of Google queries, which gives more to go on. Baird claims users are used to hopping between topics in a chat, which may make them more receptive to adjacent marketing. Koah says users click about 3 percent of ads, about on par with the more mature Google ads system.
Privacy: when asked to show an ad, Koah generally receives the user's entire conversation up to that point. People use chatbots as conversation partners and confidants, and the author says it is unclear whether users will accept the privacy trade-offs. Baird says Koah anonymizes queries and stores them in a de-identified way, but declined to comment further on data collection and storage. He argues that nonsensitive details from a chat could personalize the page that loads after an ad click, for instance filling in a feature the user asked about or a downside of a competing product the model skipped. Koah has no data on whether people want such pages. The conversation history could also support retargeting.
Negative triggers: by Koah's estimate, one in four chatbot queries about a retailer or brand concerns a return, exchange or other support issue. Baird expects manufacturers and retailers to eventually target disappointed customers with reminders about warranties and AI-run support, possibly steering them away from costlier customer service. The author notes the chatbot's own answer might help too, but buying an ad guarantees the company's message gets across. Best case, users reach a resolution faster; worst case, it is a wasted ad buy and a more irate consumer.
Agents: none of the apps Koah serves ads in offer AI agents that shop on their own. Apps such as Instinct and Meta's Muse are starting to take over shopping chores for some users. Meta CEO Mark Zuckerberg recently said Muse could become Meta's biggest source of revenue but suggested the money would come from transaction fees rather than ads. Initial studies suggest ads could still influence AI agents. Baird sees a chance to give "the user's agent the best possible recommendation," combining a great discount, a relevant product and an easy path to purchase, and says chatbots must make "abundantly clear" to the agent which content is sponsored or risk losing user trust.
The piece closes with a caution: the chatbot ad market could fizzle before agents take hold if more AI developers follow Anthropic. DeepAI CEO Kevin Baragona says ad revenue has been meaningful but he is not committed to it because the experience does not feel right yet on privacy, design and more. "I don't think the ads are working terribly well for anyone," he says.
Key facts
- OpenAI has had over 100,000 customers advertising in ChatGPT since ads debuted there in February; Anthropic keeps Claude ad-free, and its IPO filing did not mention ads.
- Google expanded ads in AI Mode to 10 more countries beyond the US last week and says users find them highly relevant.
- Koah Labs CEO Nic Baird says ads composed partly by AI generate double the clicks of fully human-made ones, and that users click about 3 percent of Koah's ads, about on par with Google.
- Koah generally receives a user's entire conversation when asked to show an ad; Baird says queries are anonymized and stored de-identified but declined to say more.
- By Koah's estimate, one in four chatbot queries about a retailer or brand concerns a return, exchange or support issue, which Baird sees as an ad opportunity.
Why it matters
Chatbots are replacing traditional search for some users, and the piece argues ads may be needed to pay for the high cost of generating AI answers. The industry has split: OpenAI sells ads to over 100,000 customers, Google is widening ads in AI Mode, and Anthropic has pledged to keep Claude ad-free. What gets decided now shapes what people see inside their AI conversations.
Who it affects
Everyday chatbot users, whose conversations may inform the ads they see. Advertisers and ad tech firms such as Koah Labs, which works with brands like Hims and UrbanStems. Niche chatbot platforms, which the author says cannot rely on the venture capital of top labs or the revenues of tech giants, and apps such as DeepAI and Liner, where Koah serves ads. Companies in regulated categories such as drugmakers also feel it, since they keep more human control over ad copy.
How to use it
For advertisers, the practical lessons come from Baird's account. Partly AI-written copy performed better, ads grounded in advertiser-supplied product details and imagery avoided hallucinated errors, and ads shown after emotionally related prompts, such as spa ads after stress prompts, did well. Baird also expects brands to target customers raising returns or support issues, and says chatbots must mark sponsored content clearly to shopping agents. For users, the piece offers no controls; it only flags that privacy trade-offs are unresolved.
How solid is it
This is a newsletter feature built on an interview with one ad tech executive. The performance numbers, double the clicks, about 3 percent click rate and one in four queries concerning support, are Koah's or Baird's own claims, and the source gives no sample size or methodology. Google's statement that users find AI Mode ads highly relevant comes without supporting data. The claim that initial studies show ads could influence AI agents is not detailed. The OpenAI, Anthropic and Google facts are stated plainly by the author.
Risks and caveats
Privacy is the main open issue: Koah generally receives a user's whole conversation, chatbots are used as confidants, and Baird declined to elaborate on data storage. Koah has no data on whether people want customized post-click pages. Dull or cluttered ads could drive users away, and the balance between engaging and risk-averse copy may be harder with no human in the loop. Agents add uncertainty: Zuckerberg suggested Muse would earn from transaction fees rather than ads. The author also says the chatbot ad market could fizzle if more developers adopt Anthropic's stance, and DeepAI's Baragona says the ads are not working well for anyone.
“I don’t think the ads are working terribly well for anyone,”
— Kevin Baragona, CEO of DeepAI