How ChatGPT Built a State-of-the-Art Ad Stack for Conversational AI

Traditional advertising systems operate on a fundamentally limited premise: they are designed to react to discrete, single-instance units of user intent. In the conventional model, a user performs a specific action—such as submitting a search query or viewing a particular webpage—and the ad system receives this event as a standalone data point. The system then functions by matching this immediate event against a database of potential ads. This process is relatively straightforward: retrieve candidates, rank them, run an auction to determine pricing, and finally display the resulting advertisement. However, the nature of conversational AI, exemplified by ChatGPT, renders this traditional approach obsolete. User intent in a chat environment does not materialize as a single, concrete request; rather, it develops organically and evolves across multiple turns, sometimes spanning entire sessions, and often without a clear, immediate 'advertising moment.'

To successfully integrate monetization into this fluid, conversational paradigm, the technical staff at ChatGPT engineered a bespoke and highly sophisticated ad delivery system. This system represents a significant departure from standard ad stacks because it is built around a custom, post-trained frontier Large Language Model (LLM). This LLM is not merely a content generator; it acts as an orchestrator, utilizing a specialized tool-call architecture to manage the entire advertising lifecycle. Its responsibilities are multifaceted, encompassing the identification of the top potential ad options (retrieval), the prediction of how engaged the user might be with an ad (ranking), the evaluation of the ad's actual usefulness in the context of the conversation (relevance), and finally, running the necessary auctions for precise ad selection and pricing.

This advanced ad stack was designed with three critical, innovative architectural decisions to support the nuance of evolving user goals. First, it implements stateful ad sessions. This means that the system does not treat every message as a fresh start. Instead, it remembers and maintains information across multiple turns within a single conversation, and even across multiple distinct conversations. This continuity is vital because user goals—such as planning a complex trip or deciding on business software—are rarely achieved in one single interaction; they are clarified and refined over time.

Second, the system relies heavily on the custom, post-trained LLM, which uses its tool-call capabilities to interpret not just the user's last message, but the entire context and the underlying, evolving intent of the conversation. This allows it to effectively select, rank, and deliver ads that are genuinely supportive of the user's journey, rather than simply reacting to the last keyword used. Third, and perhaps most critically for maintaining user trust, the entire ads selection process is run in a completely separate, independent service from ChatGPT’s core answer generation. This architectural separation is paramount; it ensures that the ad system cannot influence, bias, or compromise ChatGPT’s neutral, organic voice or the quality of the generated answer. The core ChatGPT servers handle the natural language response, while the separate ad service independently evaluates the opportunity for a sponsored placement.

The necessity of this overhaul is best understood by contrasting it with the limitations of the old request-response model. In a chat context, treating every message as a standalone request carries significant risks. These risks include misinterpreting the user's true goal by oversimplifying the context, attempting to advertise too early before the user's real need has crystallized, displaying ads that actively impede the user's natural flow, and inefficiently rerunning the entire ad flow every time the request changes, which would create noise and waste computational resources. Therefore, the ads system was built to maintain a stateful session, adapting to the developing exchange rather than relying on a single, flattened query. The LLM orchestrates the entire flow, determining the optimal ad to surface at the precise moment, based on both the user's current turn and the entire history of the conversation.

This sophisticated architecture proved incredibly successful in a remarkably short timeframe. While building a self-improving, stateful ad system like this typically requires years of development, ChatGPT Ads reached an annualized revenue run rate of $1 billion in less than 200 days after its launch. Furthermore, the system has rapidly scaled, enabling more than 100,000 advertisers to run campaigns on the platform since its debut. The combination of stateful context tracking and architectural separation is the core innovation, allowing monetization to be deeply integrated without compromising the user experience or the perceived neutrality of the AI. This technical achievement not only solves a major problem in AI monetization but also sets a new industry standard for how generative AI platforms can responsibly and effectively integrate advertising while prioritizing user trust and value.