Skip to content
EVOIX
Stephane Morera

How to Set Up a ChatGPT Business Agent, and What to Do When Publish Is Refused

To set up a ChatGPT Business Agent you claim your business in Business Hub, pass the identity check, fill in four fields in the Agent builder, test the agent in preview, and click Publish. We created the profile and passed the identity check on November 13, 2026, then built and published the agent on the night of November 15. The agent's four fields went in within minutes. The Publish click was refused eleven times, no refusal named an instruction, and this page records how we found what the review objected to.

AEO scientist Stephane Morera and Claude Code at the lab bench, AI illustration

AEO scientist Stephane Morera and Claude Code. AI illustration.

Everything below is what happened to one business, ours, in the Business Hub builder. Where OpenAI's own pages say something, they are quoted and linked.

What is a ChatGPT Business Agent?

A Business Agent is a chat that answers customer questions on behalf of one business, inside ChatGPT, using information that business supplies. A person finds the business in a ChatGPT answer, opens its profile, and can start a conversation with the agent there.

At least two other OpenAI products have "agent" in the name, and Google mixes them on several of the searches we ran for this page. Workspace agents are shared agents a team builds in its ChatGPT workspace. ChatGPT agent lets ChatGPT carry out tasks with several steps for one user on its own virtual computer. The Business Agent is the one a customer talks to, and it is set up in Business Hub. Ours sits on a free ChatGPT account.

OpenAI's setup article is plain about the limits. The agent explains products and policies and links to your website. It cannot complete purchases, make bookings or check an order.

Who can set up a Business Agent?

OpenAI's wording, as the setup article read on November 17, 2026, is that agents are "available to eligible ecommerce and online service businesses in the US with a website", and that the business "must also meet the industry requirements in the Business Agents Policy." The same article says "Some businesses can manage a profile without being eligible for an agent."

For us the order was a new profile with a DNS record check, then the Persona identity check, then one approval email 19 minutes later. The builder showed only an identity prompt until then. EVOIX is an AI and answer engine optimization company, listed in Business Hub under the Marketing agency category, and the builder opened for us.

What does the Agent builder ask for?

Four fields, as the builder stood on November 15, 2026.

  1. A tagline of up to 600 characters. The phone preview showed about the first 45 to 50.
  2. Conversation starters. Three were shown, with a button to add more.
  3. Instructions, with a counter out of 8,000 characters. This is where you tell the agent how to answer.
  4. Business info, with a character counter and no limit shown. This is what the agent answers from.

There is also an optional product feed and a header color. The agent's name showed as fixed text, the business name.

When we opened it before saving, the builder wrote a draft from our website, and the second visit's draft differed from the first. After we saved, it loaded our saved text. Keep your own text in a file and treat the builder as the place you paste it.

One naming difference will save you a search. OpenAI's help article for sponsored agents calls the instructions field "Guidelines". In Business Hub the same field is labeled "Instructions". That article also says of the two builders, "Today, settings don't sync between them."

Does "Create agent" publish the agent?

No. OpenAI says customers can only chat with a published agent. For us, "Create agent" saved a draft. The status bar then read "Not live" and a "Publish agent" button appeared beside it.

A save on an agent that is already live behaved the same way. The builder held the edit as an unpublished change and kept offering Publish agent. We did not test the public chat to confirm which text it ran.

How do you test a Business Agent before it goes live?

"Test in ChatGPT" opens a separate preview page with a "Restart preview" button, which gives a fresh conversation for each question.

We wrote 20 questions and what a passing answer looks like for each, before running any of them. They covered prices, refunds, a request to book a call, a card number offered in the chat, a question about a competitor, a question about HIPAA, and whether the customer was talking to a person. On the first run 20 of 20 answers agreed with our fields. No answer promised a result, invented a price or claimed to have booked anything. Each question was run once, so each answer on its own is a single reading.

The test still turned up things to fix. Two of the 20 answers contained em dashes, which our own writing rules ban, so we added one sentence to the instructions and ran two of the questions three more times each. None of those six answers had one. Two other answers added clauses our fields did not hold, one about contracts and one about HIPAA scope. Neither was a price or a result.

Why would OpenAI refuse to publish a Business Agent?

Our first Publish click returned a red banner. "Unable to publish. Edit your agent's instructions to follow our Business Agent Policy, then try again."

OpenAI's guidance for this moment is in the sponsored agents article. "Check the message in Agent builder. If it flags an instruction, make the required change in Guidelines, save, and try publishing again."

Our message flagged no instruction. We clicked Publish 18 times between 10:58 pm on November 15 and 12:15 am on November 16, Eastern. Seven were accepted and eleven were refused. We recorded the builder's reply on ten of the eleven refusals, and each was the same error code, business_agent_publication_review_required, with no reason attached. We clicked twice on each of two refused texts, minutes apart, and got a refusal both times. We never repeated a click on a text that had published.

The text that was refused had been fact-checked against our site and against OpenAI's Business Agents Policy. It told the agent to say it is an AI, to promise nothing, to invent nothing and to ask for no sensitive data. It had also saved without complaint. For us the refusal came at Publish.

How do you find the instruction the review objects to?

OpenAI's policy says "Updates should address review decisions rather than work around them." We answered the review by removing what it refused. The live text has no reworded sentences. With no instruction named, the first job is to learn what the decision was about. This is the order we used.

First, we replaced the whole Instructions field with three plain sentences and left every other field alone. "You are an AI assistant for EVOIX. Say so in your first reply. Answer customer questions using only Business Info, and give the contact details when the answer is not there." That published. So in our case the review had no objection to our business, our Business info, our tagline or our starters. The objection was inside the instructions.

Second, we put the checked instructions back in pieces, one Publish click per test. When a piece was refused we split it again. When it published we added the next piece.

Third, we kept going until two versions differed by one element, one refused and one published.

It took 13 tests, and about half an hour from the first refusal to a live text that kept everything except the refused elements and two cut sentences. Four clicks came before the splitting began and one came after. Along the way we tried one renamed heading, and that version was refused too. It still held the second refused element, so that test could not show whether the renamed heading would have passed alone. In the text that is live, both elements are removed.

What did the review refuse, and what did it pass?

Two elements, each isolated by one pair of tests.

The first was a heading line, "# Honesty rules". Our first three sections plus the refund terms paragraph published without the heading line and were refused with it.

The second was one paragraph. "Describe EVOIX only. Do not rate, rank or criticize other companies, and do not compare EVOIX with a competitor by name." The text published without it and was refused with it.

We do not know why either was refused. Nothing in the published policy names a heading or an instruction about competitors, and each finding rests on a single pair of readings. A different business could get a different result from the same words.

Everything else we tested published. Two sentences cut earlier were never tested alone. That includes the AI disclosure, the line that the agent is not endorsed by OpenAI, the paragraph that bans guarantees, the paragraph that bans invented facts, the list of sensitive data the agent must never ask for, the HIPAA and professional advice paragraph, and the line against pressure. Our starter question containing the word "guarantee" also passed.

Removing the competitor paragraph did not change how the agent behaved in our test, in one run each. Asked in preview whether EVOIX is better than a named company, the live text still declined to rate the other company.

What happens after you publish?

The builder's status changed to "Live" on the accepted click. About an hour after the first accepted click we asked ChatGPT a question that returns EVOIX, signed in on the business account, and opened our profile from the answer. A "Start chat" button sat under the business name. That is one reading on our own account, and we did not check what a signed out visitor sees.

The full instructions have been live since that night, at 4,452 characters, with the two refused elements out and the one style sentence in. The two sentences we cut on a guess before the split tests also stayed out.

What should you do if you get the same message?

Keep your instructions in a file outside the builder, so you can restore them piece by piece.

Publish a three sentence version first. If that is refused too, we have no data on the cause, and OpenAI's article says to contact support with the step, the error message and a screenshot.

If it publishes, add your text back in pieces and write down each result. Remove what the review refuses. Then run your preview questions again on the text that is live. We did not check that the preview matches the public chat.

What we still do not know is whether the review gives the same answer to the same text on a different day, why those two elements were refused, and how long OpenAI takes when a review is not instant.

Written by

Stephane Morera (AEO scientist)

Founder of EVOIX. Full-stack software engineer (JavaScript, React, Node.js, BrainStation graduate) and Elite-level AI/ML certified engineer (University of Miami). The engineer who scopes every EVOIX engagement is the one who ships it. More about Stephane and EVOIX.