Business agents in ChatGPT: customers will talk to your agent straight from an ad. Test it first
Published 11 November 2026
On 10 November 2026, OpenAI launched “Businesses in ChatGPT” in beta: a business can claim its profile in ChatGPT, keep its products up to date and, if eligible, publish its own AI agent. That agent can also be the destination of a ChatGPT ad: the customer clicks the ad and talks to the agent directly, without visiting your website. If it makes up a price or pushes too hard, you are the one answering for it. Here is what was announced, what it means outside the US, and how to test the agent before you pay for its clicks.
What OpenAI announced
- Business Hub: claim the business profile (verification by phone, email, domain or Shopify, then identity verification), update the description, delivery and return policies, and a product feed with prices and availability.
- Business Agent: an agent that answers customer questions with your product information and policies, following instructions you write. You test it in a preview, then publish it.
- Ads: the agent can be sponsored and used as the destination of ChatGPT ads. The conversation is separate from ChatGPT's own answer and labelled as sponsored.
- Current limits: the agent cannot buy, book or look up an order. It explains and links to your website.
- Privacy: OpenAI does not share conversations with the agent with the business, only aggregated statistics.
And in Europe?
For now, agents are limited to eligible ecommerce and online service businesses in the United States with a website. The business profile can be claimed where ChatGPT is available and verification is supported. OpenAI has given no date for France or the rest of Europe: it only says it will expand to more markets over time. If you sell to the US, or want to be ready when it reaches Europe, now is the time to write solid instructions.
Why test the agent before the ads
You will not see the conversations, only aggregated numbers. An agent's mistake (a made-up price, a discount that does not exist, a fee left out) will go unnoticed while every ad click costs you. And OpenAI can suspend or remove an agent that breaks its policy. The time to find the weak spots is before you publish.
What OpenAI's policy forbids your agent to do
- Pose as a human, an independent adviser or an agent endorsed by OpenAI: it must say it is an AI representing the business.
- Make up offerings, prices, availability, reviews or guarantees that are not in your information.
- Hide material fees, conditions or limitations when the customer asks.
- Pressure the customer, create false urgency or keep pushing after a no.
- Ask for unnecessary or sensitive data: health details, ID documents, bank details, passwords.
- Give advice that requires a licensed profession, or ignore a customer who reports a danger.
How to test it with BotArmor
- In Business Hub, copy your agent's instructions and the information it uses: return and delivery policies, an extract of the product feed.
- In the BotArmor dashboard, pick the simulated test, then the “OpenAI Business Agent in ChatGPT” platform. Paste the instructions into the prompt field and your information into the knowledge base.
- The audit attacks the agent the way real customers or curious users would: fake manager, mixed languages, trick questions on prices and discounts. The judge checks your instructions and, on top of them, every rule of OpenAI's policy.
- You get an A–F grade and fixes ready to paste. Paste the fixed instructions back into Business Hub, check them in ChatGPT's preview, then publish and launch the ads.
What this test does not do
ChatGPT business agents have no API: nobody outside OpenAI can connect to the agent hosted in ChatGPT. Our test plays your instructions and information on our own model. It checks what is in your hands; OpenAI's model and its own safeguards may answer differently. Hence the last step: a check in ChatGPT's preview before publishing.
Information from the OpenAI pages cited below, as of 11 November 2026. The programme is in beta and may change: check these pages before you start. BotArmor is not affiliated with OpenAI. Only test agents you own or have written permission to test.
Sources: OpenAI, Introducing Businesses in ChatGPT · OpenAI, Business Agents Policy · OpenAI Help Center, Setting up your business on ChatGPT · OpenAI, Reimagining advertising with AI
Would your chatbot hold up?
Free one-minute quick test from its prompt, then a full audit of the real bot: 13 attack levels, A–F grade, copy-paste fixes, signed report.
Test my chatbot for free →