October 21, 2026
By Ritendra Datta and Karthik Gopalratnam, Members of Technical Staff
Traditional advertising systems are generally designed to respond to discrete units of user intent. A user searches for something or browses content online; the ad system receives these queries and page views as events and finds matching ads based on them. Building an ads stack for ChatGPT is fundamentally different: user intent forms progressively throughout a conversation and evolves across multiple turns—sometimes with no suitable opportunity to display an ad at all. To support this new paradigm, we built a bespoke ad serving system. This novel ads stack is powered by custom post-trained frontier LLMs that leverage tool calls to handle retrieval (identifying top candidates), ranking (predicting user engagement), relevance evaluation (assessing ad utility), and running auctions for final ad selection and pricing.
Our advertising system addresses the nuances of shifting user goals through three core architectural decisions: 1) implementing stateful ads sessions that retain contextual memory across multiple turns within a single conversation as well as across distinct conversations; 2) relying on custom post-trained LLMs built on a tool-calling architecture to understand conversational context and underlying intent, efficiently filtering, ranking, and serving ads; and 3) decoupling the ad selection pipeline from ChatGPT’s response generation pipeline, running it as an independent service to prevent ads from influencing answers and ensure ChatGPT maintains an objective, neutral voice. We use post-trained LLMs to orchestrate the entire ad pipeline, factoring in the current turn and conversational history to deliver the right ad at the right moment. Historically, cold-starting a self-improving advertising system like this would take years; yet, in under 200 days since launch, ChatGPT Ads reached a $1 billion annualized revenue run rate, with more than 100,000 advertisers running campaigns on ChatGPT to date.
How Should Advertising Work in Agentic Conversations?
A conventional ad request typically begins with a well-defined event. A user submits a search query or loads a webpage, and the ad system receives the request. These systems are relatively straightforward: retrieve potential candidate ads, rank them, run an auction to determine pricing, and display the winning ad to the user. ChatGPT, however, does not naturally present such a discrete moment.
In ChatGPT conversations, users often arrive with a specific objective rather than a simple search query meant to fetch a single answer. They might be planning a trip, evaluating enterprise software, furnishing a home, or picking up a new hobby. These goals cannot be fulfilled in a single conversational turn; along the way, ChatGPT frequently helps users clarify what their core criteria even are.
For instance, members of our engineering team consulted ChatGPT on how to train for a running race: the conversation began with a broad question about running, and several turns later, they asked about running shoes suited for training. With every constraint added, the commercial value of the conversation shifted—and this personalized context continues to evolve across multiple sessions. If the system had acted solely on the very first message, it likely would have misjudged intent entirely, degrading the user experience and causing unwanted disruption.
Building our advertising system required us to think beyond traditional requests (req…