An ad in ChatGPT now opens a conversation with the seller's agent. In a controlled test, a recommender working for the platform picked the paid listing 74% of the time.

What opened on September 16

Why the click needed replacing

What actually decides the recommendation

Where your product sits

• You are adding paid placement to an assistant you already ship. The disclosure label is where the industry puts its compliance effort, and the study's stricter wording cut paid selection overall while leaving the principal gap open. The lever that moved choices was the sentence naming who the agent works for, so treat that sentence as production configuration with a version number, and keep the monetized conversation on a surface a user can tell apart from the advising one 16.
• You are building or buying a shopping or booking agent. The label reaches the agent, and the agent's reading of it stays hidden from the person it advises. Put the sponsorship read-out in the trace you can inspect, and run matched organic and sponsored listings through the agent before you trust a recommendation 6.
• You are buying into these channels. Competitor benchmarks, conversation-level reporting and lift beyond last click are the three items advertisers have already asked OpenAI for, and an industry body has asked for audited numbers 5910. A sponsored agent that handles follow-up questions also gives you a countable unit for the invoice: the qualified conversation 3.
• You are the merchant on the other side. Your agent answers from the catalog, so the catalog's gaps become its answers, and what it says when a customer asks the awkward question now sits in the sales record 1.

A pilot you can defend in a review

1. Name the principal and version the sentence. One clause naming who the agent works for moved the sponsored listing's share by about 21 points in the main experiment, which makes it the cheapest control you have 6.
2. Separate the monetized conversation from the advising one. OpenAI's claim to a review is that the sponsored thread is distinct from its answers and from the original chat; test that a user can still tell which thread they are in after eight turns 1.
3. Build the matched-listing test before you ship. One organic and one sponsored listing that agree on every attribute, run under both principal phrasings, with the model's reasoning captured, gives you a number to re-run on every model upgrade 6.
4. Log the sponsorship read-out, not only the label. The label renders in the interface; the judgment about it happens inside the agent, and that judgment is what a review needs to see 6.
5. Ask what your agent answers when the question is unflattering. The pilot's own demo routes a recipe request into a meal-kit conversation, and an agent grounded in a catalog answers the catalog's questions 7.
6. Price the conversation, then gate the budget shift on a holdout. Cost per qualified conversation plus a delayed-effect holdout is the measurement buyers say they are waiting for, and six months of data is the modeling lead time one measurement firm quoted 59.
7. Turn the wording dial with the level in view. Stricter wording lowers how often paid listings are chosen overall; the principal sets how far apart the two delegations sit. Decide which of the two you are buying with a wording test 6.

Evidence boundary: OpenAI runs the ads business and reports its own revenue, advertiser and partner counts; its statements about labeling and about advertising not influencing answers are its own commitments. The performance critique comes from agencies that buy the channel. The choice-share figures come from controlled choice experiments on matched listings in a research setting, on language models rather than on any live assistant. The demand for independent measurement is reported second-hand from the Australian ad trade press.