A New Kind of Pitch
The boundary where human judgment meets machine recommendation has rarely been so poorly marked. When a person types a worry into a chatbot — about stress, about a broken purchase, about a product that failed — the reply that comes back may look like counsel. Increasingly, it is also inventory. The advertising industry is now figuring out how to slip commercial messages into that space, and the early results suggest the seam between an answer and a sales pitch is thinner than most users would guess.
What the Numbers Show
Koah Labs, a San Francisco startup that places ads inside more than 200 AI apps, has attracted a tenth of the advertisers that OpenAI has signed up, yet its chief executive, Nic Baird, believes the company has a clear view of what works. [1] The most striking finding from Koah’s network concerns authorship: ads composed partly by AI rather than entirely by a human draw double the clicks. [1] The mechanism is not fully understood, but Baird’s working explanation is that AI-generated marketing copy likely feels more natural beside the output of an AI model. The deception here is subtle — not a false claim, but what reads as a collapse of register, a pitch that mimics the texture of advice so closely that the reader stops noticing the switch. [1]
The Regulated Exception
Not every advertiser can afford to let a model improvise. Drugmakers operate under government rules about what they must and must not say in promotions, and those companies default to heavier human control over their copy. According to Baird, that caution produces weaker results. [1]. The pattern is uncomfortable: the advertisers held to the strictest truth-telling standards are the ones whose ads perform worst, while the loosest, most machine-generated pitches win the most attention. Accuracy, in this market, appears to be a drag on performance.
Context as a Sales Tool
The machinery behind these placements is where the gap between appearance and function widens. A Google search for help with stress is likely to surface ads for therapists — a direct, literal match. Koah’s system can go further, in Baird’s view. [1]. Ads for spas have performed well when shown after prompts about stress, because the system reads not only the prompt but the AI platform’s response, drawing context from the rest of the conversation. Chatbot prompts also tend to be at least double the length of Google queries, handing the ad system more information to determine a relevant ad. [1]. What looks like a helpful suggestion is the output of a relevance engine that has been reading the whole exchange.
Familiarity as Cover
People are used to moving between topics inside a single chatbot conversation, which may make them more open to adjacent marketing than they would be in a search results page. Koah says users are clicking on about 3 percent of its ads, about on par with the more mature Google ads system. [1] That figure is the industry’s favorite kind of evidence — a number that sounds like proof of acceptance. It measures clicks, not comprehension. A person who taps an ad because it arrived mid-conversation may never have registered that a conversation had turned into a storefront.
The Privacy Trade
To choose an ad, Koah generally receives a user’s entire conversation up to that point. Baird says Koah anonymizes the queries it receives and stores them in a de-identified way, but he declined to comment any further on the data collection and storage. [1]. The admission matters because of what chatbots have become: not just search boxes but confidants, places where people work through problems they would not type into a search bar. Traditional queries ask for facts; chatbot prompts confess. It is unclear whether users will accept the privacy trade-offs that come with personalized chatbot ads as they begin to proliferate. [1].
The Promise of the Next Page
While acknowledging that personalization can become intrusive, Baird argues that there are benefits for users and advertisers alike as the tech matures. [1]. Nonsensitive information about chatbot discussions, he suggests, could inform personalized content on the page that loads after an ad click — perhaps someone expressed interest in a specific feature, or the AI model didn’t highlight a particular downside of a competing product, and the advertiser could fill in the gap. [1]. Koah doesn’t have data on whether people actually want these customized pages. [1]. The industry’s long-standing claim that relevance is welcome stands in for evidence, and the assumption does a lot of quiet work: it treats a gap in the AI’s answer as an opening for a brand rather than a failure to be corrected.
Support as a Sales Channel
By Koah’s estimate, one out of every four chatbot queries about a retailer or brand concerns a return, exchange, or other support issue. [1]. That is an opening that other channels cannot easily replicate — a chance for advertisers to address what the trade politely calls a “customer success” matter. Baird expects manufacturers and retailers will eventually target disappointed customers to remind them about warranties, engage them in AI-intermediated support, and potentially steer them away from a more costly customer service interaction. [1]. The chatbot’s own answer might offer similar help, but a paid placement guarantees the company’s message lands. The best case is a faster resolution the user can live with. The worst is a wasted buy and an angrier customer — and the user has no way to tell which case they are in.
When the Buyer Is a Bot
None of the apps Koah currently serves ads in offer users AI agents that can shop on their own, but that is changing. [1]. Apps such as Instinct and Meta’s Muse are beginning to take over shopping chores, and Mark Zuckerberg has said Muse could become the company’s largest revenue source, though he pointed to transaction fees rather than ads. Initial studies suggest ads could influence AI agents, which means the audience for a pitch may soon be software rather than a person. [1]. There is an opportunity to hand a user’s agent the best possible recommendation — a strong discount, a relevant product, an easy path to purchase — while the chatbot must make “abundantly clear” to the agent what content is sponsored, or risk losing trust. The safeguard rests on disclosure that the end user may never directly see.
The Dissent
Not everyone in the business is convinced the model holds. Kevin Baragona, chief executive of DeepAI, says ad revenue has been meaningful but he is not committed to it, because the experience does not yet feel right on privacy, design, and other grounds. His verdict is blunt: “I don’t think the ads are working terribly well for anyone.” That skepticism lands hardest against the optimism of the ad-tech firms, because it comes from someone already running the placements.
The Holdout
Anthropic is holding firm on the pledge made in its Super Bowl commercial this year to keep Claude ad-free, and its recent IPO filing did not even mention the possibility of selling ads to make money from the AI chatbot. [1]. The stance is easier to hold with about $112 billion in investor cash behind the company. [1]. For smaller platforms without that cushion, advertising may look less like a choice than a survival requirement. The result is a split market: one tier of assistants that can afford to stay clean, and a wider tier where the answer and the advertisement arrive in the same breath. A user moving between them has no reliable way to know which kind of reply they are reading.
What the User Is Left With
The people affected are heard last, and what they describe is not outrage but uncertainty. A person who asks a chatbot about stress and receives a spa recommendation cannot tell whether the model understood the problem or the ad system did. A shopper mid-return has no reliable way to know whether the warranty reminder is help or a nudge away from a human representative. Baragona’s doubt and Baird’s confidence both point at the same unresolved question, and neither settles it. The technology has learned to sound like a confidant. The commercial layer underneath has learned to sound like the technology. Until someone draws a line that users can actually see, the burden of telling one from the other falls on the person typing — who was, after all, only asking for help.