What changed
OpenAI announced the development of a bespoke advertising delivery system for ChatGPT, designed to handle the unique nature of conversation-based advertising. Unlike traditional ad systems that respond to discrete user events, this new stack supports evolving user intent across multiple conversation turns. The system uses a custom, post-trained large language model (LLM) to interpret conversation context, select, rank, and deliver ads while maintaining ChatGPT’s neutral voice.
What this means in practice
The new ad stack supports a stateful session, meaning it remembers conversation history across turns and multiple sessions. This allows the system to adapt to changing user goals, such as planning a trip or learning a new hobby, without misreading intent. Ads are selected and delivered through a separate service from ChatGPT’s organic response generation, ensuring ads do not influence the chatbot’s answers. The system also filters ads based on quality and relevance, only showing ads that are clearly beneficial to both users and advertisers.
Limitations and availability
The ad system operates independently of ChatGPT’s core functionality, meaning it cannot delay or disrupt the chatbot’s performance. OpenAI reports that the system reached a $1 billion annualized revenue run rate within 200 days of launch, with over 100,000 advertisers having run ads on ChatGPT since its introduction. The system relies on a post-trained LLM orchestrator to interpret conversations and make tool calls for retrieval, ranking, and delivery. The exact details of the LLM’s training and performance are not specified in the announcement.
