The Maze: OpenAI is testing an ad that does not send the click to a website. It opens a branded AI conversation inside ChatGPT. Wayfair has confirmed a limited Sponsored Agent pilot, with guardrails for product accuracy, transparency and handoffs to service staff. The sales pitch is obvious: answer a shopper's specific question before a generic landing page loses them. The tradeoff is sharper. The retailer rents not only the discovery surface, but also the first commercial conversation.
The destination has changed. A standard ChatGPT ad sits below a response with a headline, image and landing-page link. The new format, shown to selected clients and verified through a video of the experience, adds a call to action such as `Chat with us`. The click opens a separate branded conversation with the advertiser's AI representative. OpenAI has not disclosed how many advertisers or users can access it, where the pilot runs, or when it might expand. This is a format test, not a broad launch.
The agent turns product data into ad creative. OpenAI's terms define Sponsored Agents as advertiser-backed conversational experiences. The input can include catalog data, inventory, availability, pricing, promotions, website content, business logic and service scripts. That makes the ad dynamic: instead of pre-writing every answer on a landing page, a brand can respond to the shopper's exact question. But the advertiser remains responsible for the agent's configuration, actions and output. A stale price or invented delivery promise is not merely a model glitch. It is an advertising and customer-experience failure owned by the brand.
Wayfair is testing the handoff, not claiming a conversion win. The home retailer says its limited pilot includes guardrails around product accuracy, transparency and escalation to service teams. Those are the right pressure points. Furniture purchases produce questions about dimensions, materials, compatibility, delivery and returns. A good agent could qualify intent and reserve human help for harder cases. A bad one adds another layer of friction or misinformation. No public evidence yet shows the format improves conversion, reduces service cost or produces a positive return on ad spend.
Measurement gets harder when the click stays inside the platform. OpenAI's current advertiser guide describes cost-per-click and cost-per-thousand-impression buying, a relevance-weighted auction and reporting for impressions, clicks, spend and conversions. It does not say whether Sponsored Agents use the same pricing or which events they expose. Operators will need to separate an agent open, a useful exchange, a service handoff, a site visit and a completed purchase. Chat volume is activity, not incrementality. The privacy line also needs to stay visible: OpenAI says advertisers cannot see private ChatGPT history, but can see messages a user deliberately sends to them through an ad.
Why it matters: Google trained advertisers to buy intent and Meta trained them to buy attention. OpenAI is testing whether brands will buy a conversation. That could remove landing-page friction and make product education more responsive. It could also move qualification, merchandising and part of the customer-data trail into a platform the retailer does not own. The winning metric is not cost per chat. It is incremental margin after agent errors, service handoffs, lost site traffic and platform fees are counted. Until pricing and conversion evidence arrive, Sponsored Agents belong in controlled test budgets, not channel forecasts.
Sources: Digiday | OpenAI Ad Tools Terms | OpenAI Ads Basics | OpenAI Ads FAQ | OpenAI Ad Policies
