OpenAI made the ad a chatbot — and the thing it promises to keep separate isn't the thing that moves the decision

What actually shipped

The format is simple enough to describe in a sentence. A ChatGPT user on an ad-supported tier sees a sponsored placement. Clicking it does not navigate anywhere. It opens a second, visually distinct, labelled conversation with an AI agent the advertiser controls — one that answers questions about that advertiser’s products and, when the user is ready, hands off to the merchant’s own flow.

OpenAI’s example is furniture: asking the sponsor how a dining table fits a room, how many people it seats, how to care for the finish. Wayfair confirmed participation at limited scale, with stated guardrails on product accuracy, transparency and handoffs to customer service. Angi’s own announcement is more concrete about the endpoint — a homeowner discusses a project with the Angi agent and then enters Angi’s service-request flow to be matched with local pros. Angi co-founder Angie Hicks framed it as going “directly from discussing a home project in ChatGPT to connecting with a skilled local pro,” and noted the distinction from Angi’s existing ChatGPT app, which a homeowner has to find and add themselves.

That last detail is the whole strategic point, stated by the advertiser rather than the platform. The app required intent. The Sponsored Agent requires a click on something placed in front of you.

Around the format, OpenAI shipped the machinery of an ad business. Inside Ads Manager, campaigns can now be created and analysed by prompt, with a confirmation step before anything executes. Ad copy and imagery are drafted automatically from the landing page and the campaign objective, subject to advertiser review. An opt-in setting adapts headlines to the conversation they appear in and translates them into the reader’s language. HubSpot became the first CRM partner — ads created, tracked and followed up inside HubSpot, with automatic UTM tagging that links clicks to contact records, launched with bundled discounts and a $750 spend match through 30 September. Shopify became the first ecommerce partner; its free app has been live for US merchants since 3 September, syncing catalogues and installing conversion pixels, with international markets from 23 September. Measurement got the unglamorous upgrades that signal seriousness: impression-billed conversion optimisation generally available, configurable 7-, 14- and 30-day click attribution windows, five platform-targeting categories, and an Event Quality Score diagnostic.

No pricing for Sponsored Agents was published. No transcript-handling policy was published either.

The separation is real. It is also the wrong boundary to watch.

Give OpenAI its due: the guarantee it wrote down is the correct one, and it is stated plainly. “The conversation with a Sponsored Agent is distinct from ChatGPT’s independent answers and separate from the original conversation that the user started in ChatGPT.” Two walls — the sponsored agent does not contaminate the organic answer, and it does not get handed the session that preceded it. Both of those were the obvious failure modes, and both are addressed.

The problem is that neither wall touches what actually determines a purchase.

When ads arrived on ChatGPT’s free tiers, the format was a contextual card below the answer. A card is a thing you read and dismiss; the answer above it stayed the centre of gravity. A Sponsored Agent is not a card. It is the same interaction primitive as the product itself — a text box, a patient interlocutor, memory within the thread, follow-up questions, apparent expertise. It is separated from ChatGPT’s answer by a label and a visual boundary, and it is continuous with ChatGPT’s answer in every property a user actually reasons with.

This is why the interesting risk is not a corrupted recommendation. It is asymmetric presence. Say the organic answer names four credible options in a category. One of those four has bought a Sponsored Agent. The answer is untouched, the label is honest, nothing is hidden — and one vendor now gets fifteen minutes of consultative conversation while the other three get a name in a list. No manipulation is required for that to change outcomes. The February ad test was contextual and OpenAI said it did not influence answers; those claims can stay true while the decision moves anyway.

And one shipped feature does cut across the boundary in a way worth stating precisely. The opt-in that rewrites headlines to fit the conversation and translates them into the reader’s language means conversation context reaches ad copy — not the advertiser, but the creative. The answer is walled off. The frame is being written from what you just said.

The gap: what comes out of the sponsored transcript

OpenAI described what a Sponsored Agent does not get. It did not describe what an advertiser gets.

Inside a sponsored conversation a user will plausibly disclose room dimensions, a renovation budget, a move-in date, a household composition, a problem with an existing product. That is first-party data of unusually high quality, volunteered in a setting the user experiences as an extension of an assistant they already trust. Who retains it, for how long, whether it joins an advertiser’s CRM — the HubSpot integration makes that question concrete rather than hypothetical — and whether it feeds targeting are all unanswered as of publication.

That is not an accusation. It is an unwritten policy on a live product, and the honest thing to do is flag it as the item to watch rather than assume the worst. The reasonable user posture in the meantime is simple: a Sponsored Agent is a vendor’s sales system wearing the interface of your assistant. Say to it what you would say in a live chat on the vendor’s website, and nothing more.

Why this is a US-only test

The geography is doing work. OpenAI’s ordinary ChatGPT ads reached the EEA and Switzerland in August under OpenAI Ireland, dragging along an unresolved consent-or-pay problem. Sponsored Agents would land on top of two further constraints.

First, an advertiser-operated agent talking to a person is an AI system interacting with a natural person, which puts it inside the transparency duties that became applicable under Article 50 of the EU AI Act on 2 August 2026. “Clearly labelled” is the right instinct, but at platform scale the label becomes a compliance system that OpenAI would be operating on behalf of every advertiser in the programme. Second, ChatGPT’s designation as a very large online search engine under the DSA brings ad-repository and profiling obligations that a conversational ad unit has no established template for.

Read those together and “testing with select advertisers in the United States” looks less like a phased rollout and more like the only jurisdiction where the format is currently shippable.

What to do

If you use ChatGPT free to research purchases. Change one habit: when a sponsored conversation is genuinely useful — and for Angi’s use case it plausibly is — end it by taking the same question back to the unsponsored assistant, or to a second tool entirely. Perplexity versus ChatGPT is a live comparison precisely because citation-first and answer-first products fail differently, and a category where one vendor has bought the follow-up conversation is the moment a second opinion is cheapest. The ChatGPT review has the current tier breakdown if the ad-free line is what you are weighing.

If you buy AI tools for a team. Assume that within two quarters “which AI tool should we use” is a query with a sponsored agent attached, run by one of the vendors in your shortlist. That does not make the assistant untrustworthy; it makes single-source shortlisting worse than it already was. Our AI chatbot rankings and Claude versus ChatGPT exist to be a second input, and the value of any second input rises the moment the first one becomes monetisable.

If you sell something. This is a real channel for considered, describe-it-first purchases and a waste of money for anything bought by name. The infrastructure argument is stronger than the format argument right now: catalogue sync via Shopify is free and already live, HubSpot attribution is automatic, and the 23 September international date means the decision window for non-US merchants is this month. Our marketers and small business guides cover the surrounding stack; treat a Sponsored Agent as a support surface you now own, because an agent that misdescribes your product is your problem, not OpenAI’s.

If you are watching the strategy. File this next to Anthropic’s decision a day earlier to fold agentic work into the default chat box. Two vendors, two days, one direction: the conversation is becoming the only surface, and everything else — documents, actions, now advertising — is being routed into it. The consolidation that makes these products easier to use is the same consolidation that makes the box harder to reason about. Every additional thing that can speak in that box arrives with a label and without a habit for reading it.

Sponsored Agents are labelled, separated, opt-in at the click, and currently confined to one country. None of that is in dispute. What OpenAI built is an ad that can do the one thing an ad has never been able to do — answer a follow-up question — and the honest version of that sentence is that it is a better product for advertisers and a harder environment for buyers, at the same time.

Update, 21 September 2026 — the measurement layer under the format. This piece asked what advertisers receive from a sponsored conversation and noted OpenAI had published nothing about it. A 20 September technical investigation answers the adjacent question — what the measurement side collects on the advertiser’s own website — and the answer is more than the conversion events an advertiser configures. OpenAI’s pixel SDK draws identity from four channels including form-field scraping, rendered page text and interception of the tag-manager bus, with scraped identity outnumbering advertiser-supplied identity 685 events to 255, and automatic matching enabled on 638 of 881 pixels where the setting was observable. Carrying it home is a __obi cookie on .openai.com, one-year, SameSite=none, account-bound — and classified by OpenAI as an analytics cookie, so refusing marketing consent does not stop it. For the merchants this article told to decide on the Shopify and HubSpot integrations this month, that is the second half of the decision. The mechanism, the numbers, and the GDPR exposure that lands on the site owner.

Update, 22 September 2026 — the other half of a commerce surface. This article covered advertising that acts. On 21 September OpenAI added the targeting-adjacent data: an authorised Experian credit-report connection for US Plus and Pro subscribers, refreshing a VantageScore 3.0 and the underlying report monthly, with alerts on new inquiries, accounts and addresses. Experian has spent 2026 building the matching commerce layer inside ChatGPT — an insurance marketplace in February, a personal loans app in June, and a Credit Cards app on 27 August that surfaces partner offers matched to a user’s Experian credit profile. A Privacy Center shipped the same day and changes no existing setting. The credit-file connection in full.

Questions or corrections? Email Pick Right. Want the full list? See all news.