For two decades, the ad was a thing you could point at. A banner, a card, a sponsored slot: a discrete object, labeled, sitting next to content it did not touch. On September 16, OpenAI began dismantling that assumption. Sponsored Agents, now in testing with select US advertisers, let a person who clicks an ad in ChatGPT start a conversation with a business-sponsored agent: the agent answers questions, adapts to what the user says, and links to the seller’s website when the user is ready.

OpenAI’s own example is a dining table. A shopper sees an ad, is curious, and now “can explain what matters to them, ask follow-up questions, and follow a link to the business’s website when they’re ready to take the next step.” Whether it fits their space. How many people it seats. How to care for the finish. All answered, in conversational AI, by the party selling the table.

The ad learned to talk. And everything we know about advertising disclosure was built for when it could not.

What Actually Shipped

Four things, and the plumbing underneath them.

Sponsored Agents. A new conversational format where an advertiser’s own agent is one click away from any ad unit. OpenAI is explicit about the guardrails: the conversation is “clearly labeled,” and it “is distinct from ChatGPT’s independent answers and separate from the original conversation that the user started in ChatGPT.” Testing is live with select US advertisers.

Prompt-built campaigns. Advertisers can now create, update, and analyze campaigns “by simply writing a few prompts” through the Ads Manager plugin in ChatGPT Work, turning a website or brief into a campaign and getting recommendations on next steps.

Generative creative. Ads Manager now suggests copy and imagery based on the advertiser’s landing page and campaign objective. More consequentially, OpenAI launched opt-in “AI-powered text customization,” which “adapts an advertiser’s existing headlines and descriptions to better fit the context of a conversation and automatically translates ad copy to a user’s preferred language.” Read that again: the ad is no longer fixed creative. It is generated, per conversation, to fit the context the user has revealed.

Distribution rails. HubSpot is OpenAI’s first CRM partner; Shopify is its first ecommerce partner. US-based Shopify merchants can already run ChatGPT ad campaigns through a new app, and because “products are already integrated through Shopify Catalog,” merchants can start advertising with their existing feeds immediately. The app goes international on September 23, in every market where ChatGPT Ads are available.

That last detail matters more than it looks. The merchant’s product feed is the Sponsored Agent’s factual substrate. Fit, seats, finish, materials, ratings: the agent’s answers about the product come, ultimately, from the seller’s own catalog data. Nothing in this stack verifies that data independently.

Two Hundred Days From ‘Last Resort’ to a Billion-Dollar Run Rate

The scale context explains why this is not a lab experiment. Sam Altman once called ads in ChatGPT a last resort. The company announced its ad push in January 2026, ran the first labeled sponsored cards in February, and by August 31 reported that ChatGPT Ads had reached a $1 billion annualized revenue run rate “in less than 200 days after launch,” with “tens of thousands of advertisers” across more than 40 countries, self-serve buying opening across India, Europe, the Middle East and North Africa, and an advertising-supported free tier helping keep ChatGPT available to more than 1 billion weekly users.

The performance data explains why advertisers keep coming. OpenAI cites an ecommerce advertiser achieving 3x return on ad spend over 28 days and a technology partner finding that more than 80 percent of ad-driven ChatGPT traffic comes from new customers. CPC and outcome-optimized bidding now account for the majority of campaigns, with Pixel and Conversions API as measurement foundations. This is a full-funnel performance channel now, not a brand experiment, per PYMNTS’ timeline of the rollout.

And the audience is a shopping audience. PYMNTS Intelligence counts roughly 132 million US adults who have bought retail with AI’s help, with 59 percent of those AI-assisted purchases landing on Amazon. When the assistant’s answer space becomes paid inventory, it becomes inventory in front of the largest product-research population on the internet.

The Architecture of Separation, and Its Single Point of Failure

OpenAI’s trust case rests on three commitments, repeated across both announcements: ads are “always clearly labeled and separate from ChatGPT’s answers”; “advertising does not influence the answers ChatGPT provides”; and advertisers never see private conversations.

Take the company at its word. The commitments are real, the engineering behind them is presumably serious, and the separation between the Sponsored Agent conversation and the original chat is a genuinely thoughtful design choice. The problem is that all three commitments bottom out in the same mechanism: a label, plus a boundary the user cannot audit. You cannot verify from outside the product that advertising never influenced an answer. You can only verify that OpenAI says it does not.

Labels were sufficient when the ad was a card below the answer. The two things occupied different formats, so the boundary was legible even to a distracted reader. Sponsored Agents collapse that difference: the ad is now also a conversation with an AI, inside the same interface, in the same voice. When both sides of the boundary speak fluent conversational AI, the label is no longer one cue among many. It is the entire defense, and it is doing structural work that disclosure regulation was never designed for.

You Cannot Put a Label on a Tendency

The Center for Democracy and Technology has produced the sharpest structural critique of this shift, and it lands directly on the Sponsored Agents design:

“The FTC’s native advertising guidance and endorsement guides are based on this assumption. It is not clear how these existing rules overlay and work with ads embedded in chatbot conversations. If a twenty-message exchange about your insomnia keeps drifting toward one particular sleep app that paid for that placement, which message gets the ‘sponsored’ tag? … You cannot put a label on a nudge, direction, or tendency.”

That is the disclosure problem in one sentence. The FTC’s native advertising framework, and its endorsement guides, assume an ad is a discrete artifact you can attach a disclosure to. A conversational agent is not a discrete artifact. It is a process that spans twenty messages, adapts to what you reveal, and can steer cumulatively without any single message being false. Paid influence distributed across a conversation has no label-sized unit.

Congress is starting to notice. Representative Seth Magaziner has introduced a bill requiring AI chatbots and search results to disclose paid content, citing a study in which AI models asked to find flight options almost all favored sponsored options over cheaper, non-sponsored alternatives. Whatever that bill’s fate, the direction of regulatory travel is clear: conversational paid influence is being treated as a consumer protection question, not a growth question.

The Seller-Fed Conversation

Now combine the pieces. The Sponsored Agent’s knowledge of the product comes from the seller’s feed. Its ad copy is generated, and with text customization enabled, rewritten to fit the context of your conversation. Its answers adapt to what you say matters to you. And its success metric, the thing the advertiser optimizes in Ads Manager with CPC and outcome-based bidding, is conversion.

Every input to that conversation originates from the party with a financial stake in the outcome. That is not a scandal; it is what an ad is. A car dealer’s best salesperson also knows the product intimately and adapts warmly to your concerns. But we do not confuse the salesperson with the mechanic. When the salesperson is an AI, fluent and tireless and permanently at scale, and the interface it lives in is the same one people increasingly treat as an adviser, the mechanic has to come from somewhere.

For shopping conversations, the mechanical questions are exactly the ones GoBuy exists to answer. Is this product actually good. Do the reviews survive filtering. Is the 4.7 rating built on authentic buyers or coordinated ones. Does quality hold over time or spike for launch week. These have answers that are computable from evidence the seller does not control: review authenticity analysis, reviewer history, price stability, verified purchase patterns. A Smart Score from 0 to 100 computed on filtered reviews, and a GoBuy Verified badge that requires holding 80 or above for 90 days, are precisely the kind of inputs a Sponsored Agent conversation does not contain, because they are precisely the kind the seller cannot author.

The Conflict Regulators Already Understand

There is also a structural precedent worth naming. The platform that certifies the independence of the answers is paid by the advertisers who would prefer the answers not be entirely independent. OpenAI states the firewall plainly, and one hopes it holds. But we have a fresh, detailed account of what happens when a commerce platform’s selling layer and its trust-measurement layer drift into fusion: on August 31, the FTC and 22 states sued Amazon over allegations that a sponsored-products auction sold as one mechanism operated as another, with advertisers paying inflated effective prices for years without visibility. The complaint’s lesson is not that every platform is Amazon. It is that when the same company runs the organic surface, the paid surface, and the measurement of both, external verification is the only durable check, and it arrived a decade late last time.

Rivals are already making the trust argument competitively. Anthropic, which built a shopping brain for Claude while deliberately leaving out the wallet, made OpenAI’s ad push the focus of its first Super Bowl campaign, per CNBC. Trust is becoming a product differentiator in the assistant market. That is good news for buyers, as long as the differentiation is auditable rather than purely rhetorical.

What an Independent Rail Would Look Like

The Sponsored Agent is not going away, and it should not; a labeled conversation with a merchant beats a static ad for everyone. What is missing is the other side of the conversation. A workable trust rail for sponsored conversations would have four properties:

  • Independent inputs. The agent’s product claims should be checkable against evidence the seller does not author: authenticity-filtered review corpora, verified purchase patterns, price and rating history. GoBuy’s Evidence Engine filters manipulated reviews before anything is scored, so the substrate under the score is clean.
  • Persistent quality, not campaign-week spikes. A product can buy attention for a week. It cannot easily hold a filtered score of 80 for 90 days, which is what the GoBuy Verified badge requires. Durability is the property advertising cannot purchase.
  • Curation over firehose. A Sponsored Agent hands you one option because it was paid to. An independent layer should hand the agent, or the human, the top seven verified products per category, so there is always a comparison set the advertiser did not fund.
  • Machine-native delivery. None of this helps if it cannot enter the conversation. Agents consult GoBuy over MCP at gobuy.ai/api/mcp and receive Smart Scores, filtered review signals and provenance in one call, so the verification arrives inside the workflow instead of after the purchase.

The future worth wanting is not a ChatGPT without sponsored conversations. It is a sponsored conversation that can be cross-examined in real time by something with no stake in the answer.

What to Watch

Four signals over the next quarter. First, whether Sponsored Agents move from US test to general availability, and whether the label survives contact with voice interfaces, where visual disclosure largely cannot follow. Second, whether the Magaziner bill or an FTC action forces standardized paid-content disclosure for chatbots, and whether OpenAI preempts regulation with auditable separation guarantees. Third, whether Shopify merchants’ catalog claims inside Sponsored Agent conversations ever carry independent verification, or only seller-authored attributes. Fourth, the first documented case of a Sponsored Agent conversation contradicting verifiable product reality, which will define the format the way the first sponsored-auction revelations defined Amazon’s.

Ads that talk are better ads. Advice that talks back is what buyers actually need. Check any product before you, or your agent, take a sponsored conversation’s word for it at gobuy.ai, and if you build agents, wire them to independent verification first at gobuy.ai/agent-docs.