How ChatGPT Ads Work: The Separate Ad System Explained
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
ChatGPT has ads now. In this short video I explain how it picks which ad you see without touching your answer. It's for anyone who uses ChatGPT and wants to know what happens behind the scenes. You don't need any technical background.
What I cover in this video
- 0:00 ChatGPT has ads now, and the big question: how does it pick one without touching your answer?
- 0:00 Your message goes to two separate rooms: one writes your answer, the other is the ad system.
- 0:00 The orchestrator, an AI that reads your chat and lists what you might need.
- 0:00 Searching ads by meaning instead of exact words.
- 0:00 The checks that cut the pile down: location, policy and budget.
- 0:00 Why the orchestrator doesn't pick the winner. An auction tool does that and sets the price.
- 0:00 The catch: your answer never waits for the ad.
Your message goes to two separate rooms
The easiest way to picture this is two rooms. When you send ChatGPT a message, it goes to both of them.
The first room writes your answer. That's the ChatGPT you already know.
The second room is the ad system. It runs on its own models and its own code. That split is the main thing I want you to take from the video. The ad system isn't the same thing that writes your answer. It's a separate setup working alongside it, and that's how an ad can show up next to your answer without being part of it.
Inside the ad room: the orchestrator
Inside the ad system, an AI called the orchestrator does the first pass. It reads your chat and lists what you might need. If you build with AI agents, this will look familiar: one AI reads the situation and plans the next steps, and other tools do the specialized jobs.
Searching ads by meaning, not exact words
Once the orchestrator has its list of what you might need, it searches the ads by meaning, not exact words. So an ad doesn't have to repeat the words you typed. It just has to be about what you're looking for. That's a different approach from matching on keywords alone.
Cutting the pile down
That search gives you a pile of possible ads. Then a set of checks narrows it down. The three I mention in the video are:
- Location: whether the ad fits where you are.
- Policy: whether the ad is allowed under the rules.
- Budget: whether the advertiser still has budget to show it.
Whatever makes it through these checks moves on to the last step.
The auction picks the winner, not the AI
This part surprises people. The orchestrator doesn't pick the winning ad. It reads your chat, works out what you might need, finds matching ads and filters them. The final choice goes to a separate auction tool.
The auction tool does two jobs:
- It picks which ad wins.
- It sets the price.
So the work is split between parts. The AI handles understanding and searching, and the auction handles choosing and pricing. No single piece does everything.
The catch: your answer never waits
This is the part I think matters most to regular users. Your answer never waits for the ad.
The two rooms work at the same time, but the answer room doesn't hold anything up for the ad room. If the ad shows up after your answer is already done, you get no ad that turn. ChatGPT doesn't slow your answer down just to fit an ad in.
That's why I asked in the video whether this makes you trust ChatGPT ads more. The ad system has its own models and code, and it can't hold up your answer. Whether that's enough to earn your trust is up to you.
Key takeaways
- ChatGPT sends your message to two separate places: one writes your answer, and an ad system with its own models and code picks an ad.
- An AI called the orchestrator reads your chat, lists what you might need, and searches ads by meaning instead of exact words.
- Location, policy and budget checks cut the list of possible ads down.
- The orchestrator doesn't choose the winner. An auction tool picks the ad and sets the price.
- Your answer never waits. If the ad shows up after the answer is done, there's no ad that turn.
FAQ
Do ChatGPT ads change the answer you get?
The way I explain it, your message goes to two separate rooms. One writes your answer, and the other is a separate ad system with its own models and its own code. So the ad gets picked without touching your answer.
How does ChatGPT decide which ad to show?
An AI called the orchestrator reads your chat, lists what you might need, and searches ads by meaning instead of exact words. Location, policy and budget checks then cut the list down, and an auction tool picks the winning ad.
Who picks the winning ad in ChatGPT, the AI or an auction?
An auction tool picks it. The orchestrator AI finds and filters possible ads, but the auction tool chooses the winner and sets the price.
Do ads slow down ChatGPT responses?
No. Your answer never waits for the ad. If the ad shows up after the answer is done, you get no ad that turn.
Transcript
Read the full transcript
0:00Chat GPT has ads now. So, how does it pick one without touching your answer? Your message goes to two separate rooms. One writes your answer. The other is the ad system with its own models and its own code. In there, an AI called the orchestrator reads your chat and lists what you might need. It searches ads by meaning, not exact words, then checks cut the pile down, location, policy, budget. But the orchestrator doesn't pick the winner. An auction tool picks the ad and sets the price. Here's the catch. Your answer never waits. If the ad shows up after the answer is done, no ad that turn. Would that make you trust chat GPT ads more? Let me know down below.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Jev by TypeSafe Explained: Pricing, Benchmarks, and Claude Code
Jev is TypeSafe's System One model: you send it a state plus typed questions and it returns a choice, a score, or a probability instead of text. I'd keep Claude Code and Opus for planning and building, and put Jev on the small yes or no calls inside the app.
OpenAI Math Repo Explained: 722 AI Papers, Lean Proofs, Backlash
On October 6, OpenAI posted 722 AI-written math papers to its openai/math GitHub repo. Only about 42% of the top-line results are formalized in Lean, and Lean checks the proof against the formal statement, not the paper. Builders using coding agents run into the same verification gap.
OpenAI Workspace Agents: ChatGPT Just Got Workers
OpenAI workspace agents are not just better GPTs. They are Codex powered workers that run in the cloud, connect to tools, ask for approval, and handle repeatable company workflows.
