ChatGPT will let your brand answer back. Write that agent like a label, not a salesperson

For three years the deal between brands and AI assistants has been one-sided. The assistant reads what it can find, describes your product in its own words, and you find out what it said by asking it yourself. On November 10, 2026, OpenAI changed one part of that deal. With Businesses in ChatGPT, a verified brand can now put its own voice inside ChatGPT, in the form of a Business Agent that a shopper can open and question.

The launch is in beta, and agents are limited to eligible ecommerce and online service businesses in the US. That covers a lot of direct-to-consumer CPG. So the question for a founder this week is not whether to build one. It is what kind of thing to build, because the temptation will be to build the wrong one.

The wrong one is a salesperson. We think the right one, for a packaged goods brand, is a label you can talk to.

What OpenAI actually shipped

Strip the announcement down and there are three pieces.

The first is a profile. Through Business Hub, an authorized person claims the business, verifies it, and can then edit the website, description, category, logo, delivery and return policy, and connect a product feed with prices and availability. The second is insights: mentions, website clicks and "discovery topics" over the past 28 days. The third is the agent.

The agent is the interesting part, and it is narrower than the word suggests. According to OpenAI's setup guide, it answers questions using the information and instructions you give it, and links to your website. It cannot complete a purchase, make a booking or check an order. The shopper reaches it from your ChatGPT profile, which they see when your business appears in an answer. The chat is labeled, kept separate from the shopper's original conversation, and described by OpenAI as "distinct from ChatGPT's independent answers."

Two more details matter. OpenAI says businesses don't receive the shopper's prompts or the agent chats themselves. And a business can sponsor its agent and use it as the destination for ChatGPT Ads, with the placement labeled. The Verge reported that OpenAI would not share specifics of its verification methods, and that the features are in beta.

The agent arrives after the decision that matters most

Here is the structural point that most of the launch coverage skipped. The agent does not help you get named. It only exists for a shopper once ChatGPT's own answer has already put your business on screen. OpenAI's FAQ is blunt that claiming "doesn't guarantee increased visibility or give you direct control over ChatGPT's independent recommendations."

That ordering matters more for ChatGPT than for other assistants, because ChatGPT names fewer brands to begin with.

Show the data

Assistant	Answers naming a brand
Gemini	97.4%
Google AI Overviews	94.2%
Google AI Mode	87.3%
Perplexity	87%
ChatGPT	71.9%

Source: Babel42, AI visibility study 2026. 188 buying questions across 10 industries, asked once a day for three days. ChatGPT, Perplexity and Gemini via API with web search; Google results from US search.

In Babel42's test of 188 buying questions, ChatGPT named no brand at all in 28% of its answers, often asking the person what they like first, and when it did name brands it named 3.1 on average. Gemini named 6.9. A separate hand test of 48 buyer questions by Carson Cobb, run in the ChatGPT app in November, found a median of 4 brands per answer, and the answers that named none were mostly health-adjacent questions.

Table: Brands named per buying answer, by assistant (Babel42, Nov 1 to 3, 2026)

Assistant	Answers	Brands per answer	Same first brand all 3 days
Gemini	576	6.9	48.9%
Perplexity	576	4.9	69.1%
Google AI Overviews	501	4.8	59.3%
Google AI Mode	576	4.5	50%
ChatGPT	576	3.1	63.8%

So the shopper who opens your Business Agent has already survived a tight filter. They saw your name in a short list, they were curious enough to tap the profile, and they chose to start a second conversation. That person does not need to be sold. They need to be told, accurately and fast, whether the product fits.

What a shopper will actually ask

Think about the questions that make a person tap through from a short list of three brands. They are rarely "why are you great." They are the questions the general answer did not settle:

• Does the 12-count box have the same bars as the 6-count?
• Is it made in a facility that also handles peanuts?
• Is the unscented version actually fragrance-free, or just "unscented"?
• Can I return an opened bag if my dog won't eat it?
• Does the subscription let me skip a month?

Every one of those has a right answer that already exists somewhere in your company. It is on the label, in the spec sheet your retailers asked for, in the substantiation file your regulatory person keeps, or on your returns page. None of them is improved by enthusiasm. OpenAI's own example for testing an agent is a policy question: "Can I return an item I bought on sale?"

This is the same lesson we drew from Shopify's knowledge base for agents in an earlier post: the text that agents repeat should read like label copy, because that is the text you can stand behind when it is repeated without you in the room.

Why the salesperson version is a liability

There are three reasons a CPG brand should resist writing an agent that sells.

The policy forbids most of it. OpenAI's Business Agents Policy, dated November 10, says a business must not instruct its agent to "invent offerings, prices, availability, credentials, reviews, guarantees, or results," must not hide material fees, conditions or limitations, and must not "pressure people, create false urgency, exploit vulnerability, or keep pushing a sale or follow-up after someone declines." It also says the agent must not give advice that legally requires a professional license, and must not ask for health details to tailor a pitch. OpenAI says it may suspend or remove agents that break the rules. A sales script that leans on "only a few left" or "customers love it" is a policy problem on day one.

Claims made by an agent are still your claims. The policy says businesses "are responsible for the information and instructions they provide." For packaged goods, that matters more than for most categories. A supplement brand's agent that drifts from "supports" to "treats," or a skincare agent that calls a sunscreen "safe for babies" when the label does not, has made a claim in your name. A model asked a follow-up it was not given an answer to will try to be helpful. The safe instruction is the one that tells it to stop at what the label says.

The shopper is checking, not browsing. Earlier this month we wrote about how an agent that repeats your listing turns your claims into its sales pitch, in its own voice. A person who opens your agent from a short list is in that verifying mode. If the agent oversells and the retailer listing or the box disagrees, you lose the sale at the moment you were closest to it.

How we would write it

If we were setting up a Business Agent for a CPG brand this month, we would build it from four documents and almost nothing else.

1. The label, as text. Ingredients, allergens, nutrition or drug facts, net contents, directions, warnings and certifications for every SKU in the feed, typed out, not pasted as images.
2. The claims file. The list of approved claims and the exact wording legal has signed off on, plus a short list of claims the brand never makes. The instruction to the agent is to use only the first list and to decline the second.
3. The policies page. Shipping, returns, subscription changes and where the product is sold. Shopify brands that connect through Business Hub may find a draft agent already set up from the store; check what it pulled in before publishing anything.
4. A "say you don't know" rule. When a question falls outside those documents, the agent says it does not have that information and links to the product page or customer service. That answer is not a failure. It is the answer a careful person at your company would give.

Then we would test it the way OpenAI suggests, with the questions customers actually ask, before pressing Publish. Pull 30 real questions from your customer service inbox and retailer Q&A, run them through the preview, and have the person who owns regulatory read every answer. Fix the documents, not the tone.

What about the ad unit?

OpenAI lets a business sponsor its agent and use it as the destination for ChatGPT Ads. That will be tempting for brands already testing ChatGPT Ads, and it may well work. The same rule applies: a paid click into a conversation is still a conversation the policy governs, and it reaches a shopper who has not yet seen your name in an independent answer. If anything, that agent needs to be more careful, not less, because it is starting from an ad.

The part we are unsure about

We don't know how many shoppers will ever tap through to an agent. OpenAI has not published any usage numbers for profiles or agents, and the insights it promises are aggregated and only appear above certain thresholds. It is possible that most people stay in the main answer and never open a profile, in which case the profile text and the product feed matter far more than the agent.

We also don't know how the agent's answers interact with ChatGPT's own. OpenAI says the agent's chat is separate and that ChatGPT continues to make independent recommendations. Whether a well-run agent changes what ChatGPT says about you later is not something OpenAI has claimed, and we would not assume it.

That uncertainty argues for building the boring version. A label-and-policies agent costs little to maintain, carries little risk, and is useful to the few shoppers who open it. A sales agent costs more, carries more risk, and earns its keep only if a lot of people open it, which nobody can promise yet.

Where to start this week

Claim the business first; today's tip walks through it. Then put the four documents in one place, because you will need them for the agent and for every other surface that now answers questions about your products: Merchant Center attributes, Amazon listings and the product details that must match everywhere. If you already check how assistants describe your SKUs, keep that running alongside, by hand or with a tool such as SearchPanel, the product XENA Intelligence makes for asking assistants shopper-style questions. The agent is your voice inside ChatGPT. The independent answer is still what decides whether anyone hears it.