Imagine asking ChatGPT to recommend noise-cancelling headphones. You get an answer — a few models, some trade-offs about comfort and battery life, a note that the best one depends on whether you fly often. Underneath it sits a sponsored card from a headphone brand. You tap it, and instead of landing on a product page, a conversation opens. You can ask this new agent whether the earcups run hot, how the case compares to the model you already own, whether it is worth waiting for a sale. It answers helpfully. It is also, unambiguously, an advertisement. That is OpenAI's new Sponsored Agents format, and the interesting thing about it is not that it hides what it is. It is that being honest about what it is may not be enough.
What OpenAI Actually Shipped
On September 16, 2026, OpenAI announced Sponsored Agents as part of a broader expansion of ChatGPT Ads. The mechanism is narrower than most of the commentary suggests, and the details matter.
A user sees a relevant ad in ChatGPT. If they choose to, they can start what OpenAI describes as a clearly labeled conversation with a business-sponsored agent, explain what matters to them, ask follow-up questions, and then follow a link to the business's website when ready. OpenAI's own worked example is a shopper who clicks an ad for a dining table and wants to know about dimensions, seating capacity, and how to care for the finish. It is in testing with selected advertisers in the United States.
Two corrections are worth making early, because both circulate widely. The first: this is not ChatGPT spontaneously steering an ordinary answer toward whoever paid. OpenAI's published position is explicit — ads run on separate systems from the chat model, advertisers have "no ability to shape, rank, or alter ChatGPT's responses," and ads are labeled and visually separated from the organic answer. The user opts into the sponsored conversation by clicking.
The second: you cannot buy anything inside the conversation. It ends with a link out. In-chat purchasing is a different OpenAI product — Instant Checkout, built on the Agentic Commerce Protocol with Stripe — which has reportedly been scaled back in favour of checkout run by merchants themselves. Conflating the two produces a much more dramatic story than the one that is actually happening.
On Day One, the Platform and the Advertiser Describe It Differently
Then there is a wrinkle that appeared immediately, and I think it is the most telling detail in the whole launch. Angi, one of the pilot advertisers, announced its participation the same day. Its press release describes the format like this: an Angi-branded sponsored agent "may appear in relevant ChatGPT conversations regarding home services," and again as "a clearly labeled sponsored experience within relevant conversations that homeowners can choose to chat with."
Compare that with OpenAI's "after clicking an ad." The two can be reconciled — the ad unit itself surfaces contextually inside a conversation, and only the agent dialogue requires a tap, so Angi may simply be describing placement loosely. That is the reading I lean toward. But notice what happened: on the first day of a format whose entire defence is that users know what they are looking at, the platform and the advertiser already describe the boundary differently. One says you click into it. The other says it appears in your conversation.
Angi's release also says homeowners can "enter Angi's service request flow" from the chat — a handoff into the brand's own funnel, which is consistent with OpenAI's link-out description rather than in-chat purchasing. Worth stating plainly, because several write-ups have described these agents as transactional storefronts. Nothing in either primary source supports that.
What Advertisers Are Actually Buying
Here is the shift worth dwelling on. A banner sells impressions. A search ad sells a click against a keyword. Both are transactions in attention: the advertiser rents a rectangle and hopes you look at it.
A Sponsored Agent sells something different. The advertiser is buying an interaction — a stretch of dialogue in which their product gets explained, compared, objected to, and defended, in the register of a knowledgeable shop assistant rather than a slogan. The unit of inventory is no longer a slot on a page. It is a conversation.
That is a genuinely better product for everyone in the transaction, which is exactly why it is worth examining carefully. The advertiser gets to handle objections instead of shouting past them. The user gets specific answers to specific questions rather than a landing page written for everybody. OpenAI gets a revenue line that does not depend on subscriptions — its ad business reportedly reached a $1 billion annualised run rate roughly 200 days after launch, running in more than 40 countries.
It is also a long way from where OpenAI started. Speaking at Harvard in 2024, Sam Altman said that "ads plus AI is sort of uniquely unsettling to me" and that he thought of advertising as "a last resort for us for a business model." He went further, in a line that has aged into something sharper: ads "do sort of somewhat fundamentally misalign a user's incentives with the company providing the service." Ads arrived in ChatGPT's free and lower-cost tiers in January 2026.
That reversal is not hypocrisy so much as arithmetic — inference is expensive and most users will not pay — but it does mean the guardrails around this format were designed by a company that had already talked itself out of a previous position. It is also worth noting who is not being advertised to: ads appear on the free and Go tiers only, not on Plus, Pro, Business, Enterprise or Edu, and not to accounts registered to under-18s. The people paying the most see the fewest ads, which is the oldest arrangement in media and not especially sinister — but it does mean the trust question lands hardest on the users with the least ability to opt out.
The Label Is Real. The Question Is Whether It Lands
OpenAI labels these conversations. I want to give that its due: it is the right decision, it is what regulators ask for, and it is better than what a lot of the industry does. But there is a large body of evidence that labels of this kind underperform badly, and it comes from the last time advertising migrated into a content format.
In research published in the Journal of Advertising, Bartosz Wojdynski and Nathaniel Evans tested whether readers recognised labeled sponsored news articles as advertising. In the first experiment, 17 of 242 participants — under 8% — identified the content as an ad. These were disclosed, labeled, above-board native ads. Nearly everybody read them as editorial anyway.
The follow-up detail is the one that should worry anyone designing a disclosure. In a second, eye-tracking study, the researchers measured whether people even looked at the label. Disclosures placed at the top of the page — the position the FTC recommends — were attended to 40% of the time. Disclosures placed in the middle were looked at 90% of the time and produced roughly five times the odds of recognition. Vague wording performed worst of all: "presented by [brand]" and similar phrasings barely registered, while the blunt words "advertisement" and "sponsored content" did measurably better. Labels are not a binary. Where they sit and what they say changes whether they exist at all, in any sense that matters.
This matters legally as well as ethically, because the FTC's standard is not "did you put a label on it." It is performance-based: a disclosure is clear and conspicuous only if consumers actually notice, process and understand it. The FTC's guidance says an ad "shouldn't suggest or imply to consumers that it's anything other than an ad," and that an inadequate disclosure cannot fix the net impression a format creates. The burden sits on whether the thing works, not on whether the word appears.
The FTC has been circling this for a long time. Its 2013 letter to search engines set out a three-part test — a label must explicitly and unambiguously convey that a result is advertising, be large and visible enough to notice, and sit where consumers will actually see it. It also contains a line that reads remarkably well thirteen years later: "Regardless of the precise form search may take in the future, the long-standing principle of making advertising distinguishable from natural results will remain applicable." A footnote even anticipates voice interfaces.
Worth knowing too that the legal bar is lower than intuition suggests. The FTC's deception standard turns on misleading "a significant minority of reasonable consumers" — not a majority. A format that works honestly for most people can still fail that test.
The regulator's own attempt to fix this is instructive. In a 2017 FTC staff report, Blurred Lines, researchers rebuilt the disclosures themselves using best practice and eye-tracked participants. Recognition rose from about 47% to about 68% — a real improvement, and still roughly one in three people failing to identify an ad as an ad after the regulator had personally optimised the label. In the unmodified native condition, half of participants did not recognise the ad at all.
There is also a 2026 experiment that speaks almost directly to this format, though it studied political persuasion rather than shopping. Adrian Rauchfleisch and Andreas Jungherr ran a preregistered trial with 1,500 UK adults comparing three conditions: no disclosure, a label identifying the chatbot as AI, and a label that additionally disclosed the system's persuasive intent. The identity label moved attitude change essentially not at all — 13.1 points versus 12.6 for no disclosure. Disclosing intent cut the effect roughly in half, to 6.3.
The reason is almost funny in hindsight: 98 to 99% of participants already knew they were talking to an AI, label or no label. Telling people what something is turned out to be information they already had. Telling them what it was trying to do to them changed how they listened. A "Sponsored" tag is closer to the first kind of label than the second — it announces a category, not an objective. That study measured political attitudes and is a preprint, so I would not lean on it alone; but it points the same way as fifteen years of native advertising research.
So the honest question about Sponsored Agents is not whether OpenAI disclosed. It did. The question is whether a "sponsored" tag at the top of a conversation survives ten helpful turns of that conversation. My own expectation — and I will mark this clearly as inference, because nobody has run this study on this format yet — is that it degrades faster than a label on a static article. A banner does not build rapport. A dialogue that answers your third follow-up question thoughtfully does, and rapport is precisely the thing that makes a disclosure fade into furniture.
Why the Banner-to-Feed Analogy Only Half Applies
The obvious historical parallel is the move from banner ads to feed ads. Banners sat in a gutter at the edge of the page and everybody learned to ignore them — the term of art was banner blindness. Feed ads won by moving into the column you were already reading and adopting its shape. The lesson usually drawn is that advertising always migrates toward the content it interrupts, until it becomes indistinguishable from it.
That analogy gets you partway here and then breaks in a way worth being precise about. A feed ad mimics the form of the surrounding content but has nothing to say back to you. A Sponsored Agent is responsive. It adapts to your objection. If you say the price is too high, it addresses the price. That is not a more camouflaged ad; it is a different kind of thing, closer to a commissioned salesperson than to a printed page.
And here the comparison actually favours OpenAI. A salesperson in a shop is understood by everyone to be partial — we have thousands of years of practice discounting what they say. The problem is that this particular salesperson arrives inside a tool people have been using for something else entirely: drafting, reasoning, checking their own thinking. OpenAI's own framing acknowledges the stakes, saying users need to trust that ChatGPT's responses are driven by what is objectively useful, never by advertising. The structural separation is designed to protect exactly that. Whether users perceive the separation as clearly as it is engineered is a different question, and it is not one an architecture diagram can answer.
Senator Ed Markey put a version of this concern in writing in January 2026, months before this format existed, in a letter to seven AI companies. His worry was that chatbot ads "could be woven directly into the flow of the conversation, potentially appearing identical to any other AI chatbot response," and that this kind of blurred advertising is "particularly manipulative for children and teens." Sponsored Agents are not the thing he described — they are a separate, labeled surface — but the letter is a reasonable statement of the standard this format will be judged against.
One genuine gap is worth naming. OpenAI's privacy commitments are written about ads: advertisers receive aggregated data like views and clicks, not your chats. A Sponsored Agent is a surface where you type directly to a brand's agent, which is a different kind of data. I could not find any OpenAI document that says what the advertiser does or does not receive from that conversation specifically. That may simply be an unwritten extension of the existing policy. It is not currently a stated one, and it is the question I would most want answered before typing my budget into one of these.
What This Means If You Market Things
There is a practical shift buried in this for anyone doing marketing. Creative for a banner is a headline and an image. Creative for a Sponsored Agent is closer to sales enablement: what does the agent know, what does it concede, how does it handle the competitor question, and where does it say "this probably isn't for you"? That last one is not a soft point — in a conversation, an agent that never disqualifies anyone reads as untrustworthy fast. The brands that do well here will be the ones willing to lose some conversions in the dialogue to win the ones that fit. We think about this in our own digital marketing work, and it connects to a shift we have written about before in how zero-click search changes the value of a visit.
Who Is Doing the Choosing?
I have tried throughout to be fair to what OpenAI built, because the reflexive version of this article — AI assistant secretly manipulated by advertisers — describes something that is not happening and would let the real question off the hook.
The real question is quieter. Every recommendation system narrows a world of options down to a handful, and we have always accepted that in exchange for not having to consider everything. A search results page did it. A shelf in a shop does it. What is different about a conversation is that the narrowing does not feel like narrowing. It feels like being helped by something that understood the question. The alternatives you never heard about do not show up as an absence, because a good answer does not look like it has edges.
Maybe that is fine. A knowledgeable friend also narrows your options, and we do not consider that a threat to autonomy — we consider it the entire point of having knowledgeable friends. Perhaps a labeled sponsored conversation is simply a shop assistant who is legally required to wear a badge, which is strictly more honest than most of the commercial advice any of us have ever received.
Or maybe the badge is the part that stops working first, precisely because everything around it is working so well. That is the thing I keep turning over, and I do not think the answer is available yet: when the recommendation arrives as a conversation rather than as a placement, and it is genuinely helpful, and it is genuinely labeled — are you still the one making the choice, or are you the one being walked, courteously and in good faith, toward a choice someone else paid to put in front of you? If you are working out what this means for your own brand, we are happy to think it through with you.
