OpenAI started testing Sponsored Agents inside ChatGPT on September 16, 2026, and the format quietly redefines what an ad is. Click one and you do not land on a website — you open a separate, clearly labeled conversation with the advertiser’s own AI agent. I pay for ChatGPT, so I will never see a single Sponsored Agent in my own app; ads run on the Free and Go tiers only.
That gap between what I can see and what a client’s buyers will see is why I spent a morning reading the announcement, the ads help documentation, and the developer docs for the measurement pixel. This is a field report, not a review: I have never bought a ChatGPT ad, and Sponsored Agents are new enough that nobody can honestly claim a track record with them.
In this article
- What Sponsored Agents actually are, and how they differ from a banner
- Question 1: whether your client’s buyer is even in the ad pool
- Question 2: what Sponsored Agents can and cannot prove about conversions
- Question 3: who carries the risk when the ad talks back
- Question 4: what sponsored placements do to the answer around them
- Question 5: whether the channel is stable enough to bill against
Sponsored Agents are a handoff, not a banner
The unit is a conversation, and that single design choice is the whole story. According to OpenAI’s announcement of its new advertising experiences, when a relevant ad appears in ChatGPT a user can choose to enter a labeled conversation with a business-sponsored agent, ask follow-up questions, state what matters to them, and click through to the business’s site when they are ready. The sponsored conversation is kept distinct from ChatGPT’s main thread and from organic answers. Sponsored Agents are in a US-only test with selected advertisers.
The rest of the September 16 release is plumbing, and the plumbing matters more than the headline. Businesses that manage customers in HubSpot can now connect a ChatGPT Ads account and build, monitor, and follow up on campaigns without leaving HubSpot. US-based Shopify merchants can install a free ChatGPT Ads app that syncs product inventory, builds ads from the catalog, and reports spend, click-through rate, conversions, and cost per acquisition. OpenAI also shipped prompt-driven campaign management, AI-suggested copy and imagery, and an opt-in setting that rewrites headlines to fit the conversation and translates them into the reader’s language.
Read together, the message is that adopting this channel should not require rebuilding how a business already works. That is a good pitch. It is also the reason Sponsored Agents deserve five hard questions before anyone signs off on a budget.
Question 1: Is your client’s buyer even in the ad pool?
Start here, because for a chunk of B2B work the honest answer is no. Ads in ChatGPT are limited to the Free and Go tiers and shown to logged-in adults; Plus, Pro, Business, and Enterprise stay ad-free, and Free users can opt out entirely in exchange for fewer daily messages. Sponsored Agents inherit that same audience. Geography narrows it further — the test is US-only, on top of an ad product that reached roughly 41 markets by late August after launching in the US in February and expanding to the UK, Mexico, Brazil, Japan, and South Korea in August.
For a B2B SaaS founder whose buyers are senior operators expensing a paid plan, that pool skews away from the decision-maker and toward earlier-stage researchers. That is not worthless — it is top-of-funnel, and early-stage researchers become buyers. It just is not the channel you describe as “reaching your ICP where they already work.”
Targeting signals are narrower than most paid channels too. In Europe, ads are selected on chat topic, general location, and device type only. In the US, if ads personalization is on, ChatGPT may reference memories and recent chats when picking an ad. Advertisers do not receive chats, chat history, memories, names, emails, precise location, or IP address.
Question 2: What can the Sponsored Agents pixel actually prove?
Less than a client will assume, and the developer documentation is refreshingly specific about it. OpenAI’s ads measurement pixel supports click-through attribution using configured click windows, plus view-through attribution with a fixed one-day window; where both apply, the click takes precedence. It writes an __oppref cookie that stores a landing-page reference for 30 days and an __obref cookie holding a randomly generated browser reference for 365 days. Server-side events run through a separate Conversions API.
Two details are worth flagging to anyone buying:
- Consent is a deliberate call in your code. Setting
oaiq("consent", false)before initialization blocks measurement-event pings entirely. Whoever installs the pixel decides how that interacts with the client’s existing consent banner — it is not automatic. - User data can be hashed with SHA-256 for email, phone, names, and external IDs, and the docs state raw customer information is not sent to OpenAI.
A one-day view-through window is conservative by ad-industry standards, which I read as a good sign rather than a limitation. But it means any B2B purchase with a multi-week consideration cycle will under-attribute in this channel. Set that expectation before the first invoice, not after the first report.
Question 3: Who carries the risk when the ad talks back?
Nobody has published a good answer yet, and that is the part I would not gloss over. Sponsored Agents move the exposure from creative approval to live behavior. A static ad says one thing forever. A Sponsored Agent improvises inside a live conversation, under the advertiser’s brand, in a context where the user has already been primed by ChatGPT’s own tone. The labeling is clear and the sponsored thread is separated from organic answers — OpenAI did the structural work. The behavioral work belongs to the advertiser.
A banner that goes wrong is a bad impression. A Sponsored Agent that goes wrong is a transcript, with your client’s logo on it.
I have never run an ad campaign, so treat this as an outsider’s read of an adjacent problem I do know. When I build agents for my own work, every one gets scoped instructions, a refusal boundary, and a log I actually check — the same checks I apply to any agent OpenAI puts in front of solo builders. A branding agency director I work with would call the Sponsored Agents format a brand voice surface, not a media buy, and would want approval rights over the agent’s script before the campaign goes live. That instinct sounds right to me.
Question 4: What do sponsored placements do to the answer around them?
Sponsored placements are already common enough to change how a free-tier screen reads, which affects me even on an ad-free plan. PPC Land’s roundup of the launch cites Similarweb data from August 17 indicating that 26% of ChatGPT responses carried sponsored placements. OpenAI’s position is that ads are clearly labeled, visually separated, and do not influence the answers ChatGPT gives. Those two statements can both be true and still change how a free-tier user reads the screen.
My research has not run through ChatGPT for months — it runs through Perplexity, because cited retrieval is auditable and a confident uncited answer is not. I wrote up why client research moved off ChatGPT for speed reasons before ads were part of the calculation. The arrival of Sponsored Agents does not reverse that decision; it reinforces a rule I would now state plainly: the tool you research in should not be the tool someone paid to reach you in.
For clients, the practical implication is the inverse and more optimistic. If a quarter of responses already carry a sponsored slot, the inventory exists and the surface is real. The open question is engagement quality, not availability.
Question 5: Is the channel stable enough to bill against?
The growth numbers say yes and the product maturity says wait, which is an uncomfortable but workable place to land. PPC Land cites a Digiday report from August 31 putting OpenAI’s advertising business at a $1 billion annualized run rate in under 200 days, after reaching $100 million annualized within six weeks in March. HubSpot reported 250% growth in weekly active users of its ChatGPT connector, though it did not disclose the base. Money is arriving faster than best practices are.
Against that, the calendar reads like an early access program: the Shopify app listed on September 3 with international rollout from September 23, a HubSpot spend-match qualification deadline of September 30, and Sponsored Agents themselves still a US-only test with selected advertisers. When I first wrote about ads arriving inside ChatGPT this spring, the whole thing was a test too. Tests here ship fast.
So the billing answer is specific rather than binary:
- Treat Sponsored Agents as an experiment line item, not a channel migration — a defined test budget with a defined kill date.
- Ask for access before promising performance. The test is invite-shaped, and an early-stage startup operator should not have a launch plan blocked on an approval nobody controls.
- Instrument the pixel and the consent path first. Attribution you cannot read makes the test unfalsifiable.
For me, Sponsored Agents are the first ad format in years I find genuinely interesting to think about, and still not one I would move a client’s core budget into this quarter. The audience is real but skewed young in the funnel, the measurement is honest but short-windowed, and the brand risk of an improvising agent has no established playbook. I would run a small, instrumented test with a client who has budget to learn with, and I would keep my own research on a tool nobody can buy their way into. That split has served me well enough that I am not changing it for a launch week.
Sources
- Reimagining advertising with AI — OpenAI
- Ads measurement pixel — OpenAI developer documentation
- Ads in ChatGPT — OpenAI Help Center
- OpenAI lets advertisers run ChatGPT ads from HubSpot and Shopify — PPC Land
AI-assisted research and drafting. Reviewed and published by ToolMint.