Basil C. Puglisi reads OpenAI’s new Businesses in ChatGPT feature as the directory-listing lesson with a voice attached. According to OpenAI’s November 10 announcement, a business can now claim its profile inside ChatGPT, in beta. According to OpenAI’s product page, an eligible business can then publish a Business Agent that answers customers in the brand’s name. For now, OpenAI limits agents to eligible ecommerce and online service businesses in the U.S. The listing can now talk.
The governance problem sits in what OpenAI says the business receives back. As OpenAI describes the product, feedback arrives as aggregated insights, and the agent’s conversations stay with the customer and the platform. A brand can control what its agent reads and what it’s told, and it can test the agent before publishing. It can’t later read what the agent said. Puglisi’s advice to a marketing leadership team is to put governance at the inputs and at the publish decision, with one named owner.
A claimed profile is a directory listing that can answer back
In Directory Listings 101, Puglisi told small businesses to claim the directories that matter and to match name, address and phone exactly across every listing. He also said to check monthly so nothing drifts, and he warned that wrong information in one place spreads everywhere. The same discipline applies to a ChatGPT profile.
The difference is what a mistake costs. A stale directory entry shows a wrong phone number. A stale agent can state a wrong return window to a shopper who is ready to buy. OpenAI’s product page says the agent answers using product and policy information under guidelines the business controls. It also says the person claiming must be authorized to represent the business.
OpenAI also says ChatGPT continues to make independent recommendations and comparisons, and that claiming a profile doesn’t guarantee higher placement or sales. The claim buys a say in the information and a voice. It doesn’t buy position.
The brand governs what the agent reads and can’t read what it said
OpenAI’s setup guide walks the owner through a sequence. Review the information the agent will use, such as product details and return policy. Follow OpenAI’s Business Agent policy when writing instructions. Preview the agent with a common customer question, and if the answer is wrong, update the information or instructions and test again. Customers can chat only with a published agent.
That preview is the brand’s main look at answers before customers see them. After publication, OpenAI describes aggregated insights such as mentions, website clicks and discovery topics, and says it doesn’t share users’ conversations or agent chats with businesses. The help center adds that insights cover the past 28 days and that discovery topics don’t reveal individual prompts. Those figures show discovery and engagement. They say little about whether an answer was correct.
Puglisi’s AI Visibility: Be the Human Voice AI Can Find makes the same distinction for AI answers in general. It defines AI Visibility Governance as the practice of making an organization’s knowledge findable, retrievable, citable, accurate, and accountable inside AI-mediated answer systems. It moves reporting from clicks and rank tracking to citations, answer inclusion, misattribution and verification. It also argues that a team measuring only mentions will “overcount weak visibility.” Its KPI list includes misattribution rate, answer-share accuracy, correction cycle time and source-of-truth coverage.
Machine checks help, and a named owner decides
In Puglisi’s framing, OpenAI’s verification and integrity checks, and any automated testing a team adds, are machine checking machine. That is Responsible AI: useful for catching obvious problems at speed, with no named person answering for the result. A one-off preview with no written test set and no record is a look, not a decision.
AI Governance adds a named human with accept, modify or reject authority at defined checkpoints. Puglisi’s Checkpoint-Based Governance (CBG) approach puts two of them here. The first sits before the agent is published. The second sits before every change to the product feed, the return or pricing policy, or the agent’s instructions.
The Factics sequence, Fact, Tactic, KPI, shows how this works. The Fact is that the brand can’t read the agent’s chats. The Tactic is a written test-question set and a signature. The KPI is the share of tests the agent passes under that signature.
The record matters as much as the signature. A file showing who signed, on what date, and against which test answers lets a leadership team reconstruct why the agent was cleared when a customer disputes what it said. Without that file, the only account of the agent’s behavior belongs to the platform.
Sponsorship raises the stakes. OpenAI says a business can sponsor its Business Agent and use it as the destination for ChatGPT Ads. Paid media would then land in a conversation the brand can’t read, so the signature belongs before the money goes behind it.
Six steps for marketing leadership
- Claim and verify the profile, and assign one named owner for the Business Hub.
- Treat the product feed and the return and pricing policy as the agent’s source of truth. Keep one canonical version that matches the website.
- Write the test-question set: returns, pricing, availability, comparisons with competitors, and requests that fall outside policy. Keep every answer.
- Have the owner sign accept, modify or reject before publishing and before each feed, policy or instruction change, and record the evidence behind each signature.
- Watch mentions, clicks and discovery topics, and don’t read them as proof of accuracy.
- Hold sponsorship until the agent has passed the test set under a signature.
Four measures and the limits of this advice
Puglisi would put four KPIs on the dashboard:
- Test-set pass rate before publication and after each change.
- Share of agent input changes that carry a signed accept, modify or reject record.
- Days between a website policy change and the matching hub or feed update.
- Share of discovery topics that match the services the brand actually sells.
The limits are real. OpenAI presents this as a new release with its own eligibility rules, so a business should confirm today whether it qualifies for an agent. Every product description here is OpenAI’s own. The help center also says OpenAI reviews and publishes changes to business details, and updates can take time to appear in answers. Insights are aggregated, and the business has no access to chats. Claiming doesn’t guarantee placement. CBG is also Puglisi’s working approach, in use in his own work and not independently validated at enterprise scale. Other methods may serve a given team better, and teams should test, challenge or replace this one.
The brand signs what the agent is allowed to know
Because the brand will never see what the agent said, the signature has to attach to what the agent was given. Every feed row, policy sentence and instruction becomes a statement the owner stands behind, dated and recorded. That is heavier work than filling in a profile, and it’s the work that counts. A marketing team that treats the hub as a form to complete has handed its brand voice to whatever the feed happens to say that week.
One question stays open. If OpenAI later adds transcript access or review tools, the checkpoint set should extend to the agent’s outputs. Until then, the written test set and the signed record are the only evidence a brand will hold when a customer says the agent told them something else.
Sources
- OpenAI. (n.d.). Businesses in ChatGPT: Claim your business. https://openai.com/business/businesses-in-chatgpt/
- OpenAI. (2026, November 10). Introducing Businesses in ChatGPT. https://openai.com/index/introducing-businesses-chatgpt
- OpenAI Help Center. (2026). Setting up your business on ChatGPT. https://help.openai.com/en/articles/20001635-setting-up-your-business-on-chatgpt
- Puglisi, B. C. (2025, April 10). Directory listings 101: Claim your business everywhere that actually matters. basilpuglisi.com. https://basilpuglisi.com/directory-listings-101-claim-your-business-everywhere-that-actually-matters/
- Puglisi, B. C. (2026, June 25). AI visibility: Be the human voice AI can find. basilpuglisi.com. https://basilpuglisi.com/ai-visibility-be-the-human-voice-ai-can-find/
Questions readers ask
What is a ChatGPT Business Agent?
A Business Agent is an assistant a verified business can publish inside ChatGPT to answer customers in the brand’s name. The business supplies product and policy information plus instructions. OpenAI calls it a beta, open first to eligible U.S. ecommerce and online service businesses, so confirm current terms before planning around it.
Can a business read what its Business Agent says to customers?
As OpenAI describes the product, businesses receive aggregated insights such as mentions, clicks and discovery topics. OpenAI says it doesn’t share users’ conversations or agent chats with businesses. That leaves pre-publication testing and controlled inputs as the brand’s main evidence, since transcripts aren’t available for review.
Who should own a ChatGPT business profile?
One named person with authority over the brand’s product, pricing and policy messaging should own the Business Hub. That owner holds accept, modify or reject authority before the agent is published and before each feed, policy or instruction change, and keeps a record of the evidence behind each signature.
What belongs in a Business Agent test-question set?
Write questions customers actually ask: returns, pricing, availability, comparisons with competitors, and requests that fall outside policy. Run the set before publishing and after every input change, and keep the answers. The pass rate becomes the evidence the owner reviews before signing.
Should a business sponsor its Business Agent right away?
Puglisi advises holding sponsorship until the agent has passed the written test set under a named owner’s signature. Paid traffic sent to a conversation the brand can’t read raises the cost of any wrong answer, so the publish signature matters more once money is behind the agent.
#AIg
Leave a Reply