OpenAI recently published the technical details of ChatGPT's advertising system. And no, this isn't the classic banner model we know from Google or Meta. It's something fundamentally different: ads selected by a separate LLM, integrated naturally into conversations, chosen based on the full context of a discussion, not just the last keyword typed.
A recent article on Search Engine Journal breaks down how this "ad stack" built on large language models actually works. The numbers speak for themselves: over 100,000 advertisers, one billion dollars in annualized revenue, all in less than 200 days since launch.
What does this mean for agencies and brands investing in digital advertising? Pretty much everything.
How ChatGPT's advertising system actually works
This isn't a simple ad server bolted onto a chatbot. OpenAI built what they call an LLM-based "orchestrator": a separate model that operates independently from the model generating your answer.
While ChatGPT writes your response, a second model analyzes your conversation, transforms it into search queries, converts those into embeddings (numerical representations of meaning), then searches for relevant ads in an inventory. If it finds something fitting, the ad goes through an automated auction and then through an additional relevance check, even after winning the auction.
If the process is too slow or finds nothing relevant, your answer arrives without any ad. ChatGPT doesn't make you wait to show you a placement. And perhaps most importantly: the ad doesn't influence the answer. The two systems operate independently, with separate models and code.
Ads appear labeled "sponsored" below the response, with a short AI-generated introduction that connects the ad to the conversation context. It's not a static banner. It's a dynamically constructed message based on what you've been discussing.
The selection process runs through multiple filtering stages. The orchestrator checks geographic targeting, policy restrictions, remaining budget, and campaign pacing. Candidates that pass are scored on conversation relevance, historical performance, and the likelihood of a view, click, or conversion. Only then do they enter the auction. The language model decides who participates in the auction, but doesn't pick the winner or set the price. It's a system with more filtering layers than Google Ads, and it shows in the impression rate: many conversations end without any ad, because no candidate passes all verification stages.
Why this is different from Google Ads and Meta Ads
The fundamental difference is context. Google knows what you searched for right now. Meta knows what you liked last month. ChatGPT knows what you're discussing now, what you discussed in previous sessions, and what your real intent is based on a complete dialogue.
OpenAI calls these sessions "stateful": the system doesn't restart from zero with each message. It tracks relevant details throughout the conversation and can use (for users who've opted in) non-sensitive information from previous conversations.
A concrete example from OpenAI's documentation: a user comparing project management software mentions they run a small agency, need client approvals, and want to avoid guest account fees. An advertiser offering exactly free guest access can be targeted on this complete context, not just the keyword "project management."
This is a structural shift. Traditional advertising runs on signals (clicks, likes, search queries). Conversational advertising runs on declared intent. And the difference between a signal and an intent is enormous when it comes to conversion.
For performance agencies, this means a new platform to monitor and test. For brand strategists, it means a new channel where context matters more than reach. And for brands with complex products (SaaS, B2B services, consulting), the conversational format might be more natural than any banner has ever been.
What's missing from the success story
The headline numbers are impressive: one billion dollars annualized, 100,000 advertisers. But several essential pieces of information are missing, and they should make you cautious, not skeptical.
We don't know the average CPA. We don't know the conversion rate. We don't know how many of those 100,000 advertisers return with recurring budget. There are no controlled studies demonstrating incremental sales generated by these placements. The questions around accountability in AI advertising remain largely unanswered.
OpenAI also hasn't detailed how much each advertiser sees of the performance data. The system internally logs searches with no candidates, ads that failed eligibility, and interactions with displayed placements. But it's unclear what reaches the advertiser's dashboard and what stays in internal logs.
This doesn't mean the platform doesn't work. It means it's too early to move significant budgets without testing and measuring on your own. Experiment with small budgets, compare against Google Ads and Meta Ads on the same objectives, and track real cost per acquisition, not the platform's estimates.
Another aspect worth noting: OpenAI describes the separation between answer and ad as a design decision, but the operation has been documented internally, not independently verified. As an agency, we assume platforms do what they say, but we monitor constantly. The same principle as with any new platform.
What to actually do with this information
Conversational advertising isn't a new concept. But a billion dollars in 200 days from it is. ChatGPT has demonstrated that people don't just accept ads in conversations, they interact with them, if they're relevant and don't interrupt the experience.
If you have clients with complex B2B products, services with long decision cycles, or offers where purchase intent is expressed explicitly in a conversation, ChatGPT Ads deserves a serious test. Not a 10-dollar-a-day budget, but a structured experiment with clear objectives that you can compare against other channels. Think of it like when we started exploring AI agents as a distribution channel: experimental, but with metrics front and center.
For agencies, the right time to explore isn't when the platform goes mainstream. It's now, when early adopters enjoy lower costs and less competition.
Here's what to do concretely: allocate 5-10% of your experimental budget to ChatGPT Ads for 30 days. Set up tracking correctly from the start (UTMs, offline conversions if applicable). Compare CPA and lead quality against equivalent campaigns in Google and Meta. If the numbers hold up, scale. If not, you've lost little and learned a lot. That's what we're doing ourselves, and that's what we recommend to our clients.
