News
OpenAI says ChatGPT ads use a separate LLM-driven stack to interpret intent, rank placements and keep ads separate from answers.
OpenAI said on 21 October 2026 that it has built a custom advertising stack for ChatGPT designed to interpret conversational intent, separate ads from organic answers and surface relevant sponsored placements at scale. The disclosure is significant because it describes how the company is trying to place ads inside a chatbot without letting advertising logic interfere with response quality or timing.
Evidence and context
In its engineering disclosure published on 21 October 2026, OpenAI said ChatGPT Ads uses stateful sessions and a post-trained LLM orchestrator to understand the conversation and turn it into queries and embeddings. The system then retrieves candidates through approximate-nearest-neighbor indices and calls tools for eligibility, relevance and engagement scoring. OpenAI said filters address targeting, policy, budgets, pacing and deduplication, while the orchestrator nominates candidates and an auction tool, not the LLM itself, determines final selection and pricing.
The company also said advertising runs on separate services, models and code from organic answer generation, and that slow or unavailable ad services do not delay answers. Final checks can suppress inappropriate or late ads, and the frontend shows sponsored placements below responses with advertiser creative paired with a separate model-generated neutral preamble, according to OpenAI. The disclosure did not provide exact model versions, latency targets or the auction pricing formula.
OpenAI has been laying groundwork for the format over the past few months. On 16 September 2026, it said it was reimagining advertising with AI through Sponsored Agents testing, AI-assisted campaign management and integrations with HubSpot and Shopify. On 31 August 2026, it said advertising had reached a $1 billion annualized revenue run rate and that availability had expanded to more than 40 countries; OpenAI described that figure as a run-rate metric rather than realized annual revenue.
The October disclosure said measurement and optimization rely on Pixel and Conversions API foundations. OpenAI described the implementation as a first-party engineering account rather than an independent audit, and the material does not disclose external verification of the system’s performance or user impact.
Be the first to comment