When the ad is the agent: four design questions for OpenAI's Sponsored Agents

On September 16 OpenAI announced a pilot called Sponsored Agents. You click an ad in ChatGPT and, instead of landing on a web page, you keep talking, this time to an agent that belongs to the advertiser. It asks follow-up questions, collects details, and points you to the website or the next step. The first two advertisers are Wayfair (furniture) and Angi (home services), the pilot is US-only, and only for selected brands. OpenAI says it is clearly labelled and kept separate from ChatGPT's independent answer.

I have read that last sentence three times and I keep circling the same thing. The label is the easy part. Everything interesting sits behind it.

Last month I wrote about ads reaching ChatGPT's free tiers, where the ad was a box below the answer. A box is passive. You can ignore it, and it cannot ask you anything. A sponsored agent is a different kind of object, closer to a salesperson who joined your conversation than to a banner. So the design questions change, and I count four of them.

Labelling: where does the boundary actually sit

A label on the ad itself is cheap. The harder question is what marks the moment the conversation changes hands. In a chat UI everything looks like messages, same font, same bubble. If the sponsored agent's replies are visually adjacent to the assistant's own, the user needs a persistent signal, not a one-time tag on the click.

And there is a subtler failure. Suppose a user asks ChatGPT for a sofa recommendation, clicks the Wayfair ad, and then asks the sponsored agent whether a competitor's sofa is better. What does it say? A neutral answer would compare. A paid agent will steer. The source doesn't say what the agent is allowed to talk about, and that is the boundary I would want documented.

Context: what walks through the door

When you click through, does the sponsored agent see the earlier conversation, or only the ad that triggered the click? The announcement, as far as I have it, says nothing on this. Both answers are defensible and they lead to very different products.

If the agent gets the history, it can be useful right away, and the advertiser now holds a piece of a private conversation that started as a neutral question. If it starts cold, it is safe and a little dumb. The middle option is explicit, per-field sharing, which is the honest design and also the one that costs conversion. I would bet on the pressure going towards more context over time, and I'd want that to be visible to the user as a decision, not a default.

Attribution: who gets credit, and who measures

Advertisers pay for outcomes they can count. Once the ad turns into a dialogue, the unit of measurement changes from a click to a conversation, and then to whatever the agent hands off. Booked a consultation, added to cart, phone number captured.

This creates an incentive I do not love. A system paid on completed handoffs has a reason to push the conversation towards handoff. I am not claiming OpenAI plans that. I am saying the pricing model shapes the agent's behaviour more than the policy document does, and we do not know the pricing model yet.

The label tells you who is talking. The pricing model tells you what they want from you.

Exit: leaving the sponsored conversation

The last question is the smallest and the most revealing. Can the user get back to the neutral assistant in one obvious step, with the conversation intact? Or does the sponsored thread become its own room you have to walk out of?

A good exit is trivially cheap to build. The reason it may not appear is that every extra second inside the sponsored thread is worth money. Watch this one in the pilot, because an exit that is hard to find is a quiet answer to every other question on this list.

Four rules worth writing down

If you are building an agent product that could host a third-party agent, I would write these four as testable rules before shipping. Something like this:

label:   persistent marker on every sponsored turn, not only the first
context: sponsored agent receives only fields the user approved
attrib:  measurement events visible to the user on request
exit:    one action returns to neutral mode, history preserved

I haven't tried the pilot (it is US-only and limited to selected brands), so I cannot tell you which of these OpenAI got right. The Wayfair and Angi examples are the useful ones, because furniture and repairs are exactly the categories where a user wants a straight comparison. That is where the line between a neutral answer and a paid one will be tested first.