ResearchOpenAI

OpenAI builds new advertising system for ChatGPT

Original title: How we built ChatGPT’s novel ad stack for user trust and relevance

Plain language

What it is: OpenAI built a new advertising system for ChatGPT to show relevant ads during conversations. The system is designed to understand what users want as their conversation develops over time.
How it works: The system remembers information across multiple turns of a conversation. It runs as a separate service from ChatGPT's main answer generator, ensuring that ads do not influence the actual answers you receive.
Where it is available: The system applies to the ChatGPT product.
Example: If a user asks about planning a trip to Yosemite and later asks about shoe brands, the system can connect these topics to show ads for hiking shoes.

What it is: OpenAI developed a bespoke ad stack for ChatGPT powered by a custom, post-trained frontier LLM. The system is designed to handle stateful, multi-turn conversational intent rather than single-event queries.
How it works: A post-trained LLM acts as an orchestrator, making iterative tool calls for retrieval, ranking, relevance evaluation, and auctions. The ad selection runs in an independent service with separate models and code, ensuring zero dependency from the core ChatGPT response generation.
Where it is available: The system is implemented within the ChatGPT service.
Example: The orchestrator summarizes a multi-turn conversation about a Yosemite trip and shoe brands into semantic embeddings, then queries database indices using approximate-nearest-neighbor search to retrieve relevant hiking shoe ads.

The headline and explanation on this page are based on OpenAI's post. The full details are in the original.

How we built ChatGPT’s novel ad stack for user trust and relevance