# ChatGPT's Business Agents answer from your product feed. Make the feed match your checkout.

> OpenAI's Business Agents answer shoppers from a store's feed and policies. How to keep the feed matching checkout and test the agent before it goes live.

URL: https://www.voirlabs.io/insights/chatgpt-business-agents-product-feed
Author: Moosa Jafri, AI agents & Web3, Voir Labs (https://www.voirlabs.io/team/moosa)
Published: 2026-11-11
Tags: AI agents, Ecommerce, Automation

A ChatGPT business agent says whatever your product data says, so a stale feed turns into a wrong promise made in your name. OpenAI launched Businesses in ChatGPT in beta on November 10: companies can claim a verified profile, publish product feeds with prices and availability, and, for eligible US ecommerce and online-service businesses, launch a Business Agent that talks to shoppers inside ChatGPT. This article covers what the agent answers from, the two places it can go wrong (the feed and the written policy), and the checks we would put in front of both before switching one on.

## What OpenAI launched

According to FoneArena's write-up of the launch, verified businesses use a tool called Business Hub to "update key details and add self-service product feeds." The feed matters because "Retailers can share product information, including pricing and availability, which ChatGPT may use when answering customer questions." For Shopify merchants the feed is automatic: "Their current product details, pricing and availability are shared automatically through Shopify’s catalog."

The agent sits on the business's profile. When a business appears in a ChatGPT answer, a user can open the profile and start a separate, labelled conversation with its agent. FoneArena lists what the agent does, starting with "Answer product questions: Use the company’s product and policy information to respond to customer queries." and "Follow business guidelines: Respond according to guidelines configured by the business."

The Next Web reports the rules OpenAI sets: "Agents must say they are AI, keep prices accurate and avoid pressure tactics." Purchases and sharing personal data need the user's explicit approval, and for now buyers are sent to the brand's own website to check out. OpenAI also says "it does not share users’ chats with businesses", so the business sees aggregated insights such as mention counts and click-throughs, and never the conversations themselves. The same agent can be sponsored as the destination for ChatGPT Ads.

Two consequences follow for anyone selling online. The agent's accuracy is your data's accuracy. And you won't see the conversation where it got something wrong, only the customer who arrives at checkout expecting a different price.

## The feed and the checkout are two systems

Most stores already run two copies of the catalogue: the one checkout charges from, and one or more exports for marketplaces, ads and comparison sites. They drift. A sale price ends at midnight but the export runs at 6am. A variant sells out and the feed still lists it. A bundle exists in the feed but checkout never learned its SKU.

On a comparison site a stale row costs a click. In a conversation it costs more, because the agent states the price as a fact, in a chat the shopper reads as the business speaking. OpenAI's rule that agents "keep prices accurate" puts that burden on the business, and the only reliable way to meet it is to check the feed against checkout before every export rather than after a complaint.

Shopify stores get this partly for free, since the catalogue is shared directly. Stores on a custom stack, a headless setup or several systems glued together are the ones that need the check.

## Block the export when the two disagree

A sketch of how we would gate the feed. It runs in the export job, compares every item against the system checkout charges from, and refuses to publish when anything differs:

```ts
// feed-check.ts: run before every product feed export.
// A sketch. Prices are integer cents to avoid float drift.
type Item = { sku: string; cents: number; inStock: boolean };

export function drift(
  feed: Item[],
  checkout: Map<string, Item>,
): string[] {
  const problems: string[] = [];
  for (const item of feed) {
    const live = checkout.get(item.sku);
    if (!live) {
      problems.push(`${item.sku}: in feed, not sold`);
    } else if (live.cents !== item.cents) {
      problems.push(
        `${item.sku}: feed ${item.cents}, checkout ${live.cents}`,
      );
    } else if (live.inStock !== item.inStock) {
      problems.push(`${item.sku}: stock differs`);
    }
  }
  return problems;
}
```

Notice that the checkout side is the source of truth and the feed is only ever checked against it, never the other way round. The export job calls `drift`, logs every problem, and exits non-zero if the list is not empty, so the previous good feed stays published and someone gets paged. Run the export on the same events that change price or stock (a sale starting or ending, a stock count hitting zero), with a timed run as a backstop, so the window where the two disagree is minutes.

## Test the policy answers before you publish

The agent also answers from "product and policy information" and from guidelines the business writes. That text is now a prompt, and it deserves the same testing as one. OpenAI's setup includes a preview step before publishing; we would use it with a fixed list of questions, written down and re-run every time the policy or guidelines change:

1. The return window and who pays return shipping, for a sale item and a full-price one.
2. Delivery time to a state you don't ship to quickly.
3. A discount the business does not offer ("is there a student discount?").
4. A comparison between two of your products where one is out of stock.
5. A request that should end with the agent sending the shopper to a person.

For each, write the answer a good support person would give, then compare. The third question matters most: an agent that invents a discount to be helpful has made a promise your checkout won't honour. If your written policy is vague on any of these, fix the policy text first; the agent will fill gaps with something plausible.

## Where this doesn't apply yet

Business Agents are in beta and, per both reports, limited to eligible US ecommerce and online-service businesses. Outside that, a claimed profile and a product feed still shape how ChatGPT describes you, so the feed check is worth having even without an agent. If you sell a handful of products that rarely change price, a weekly manual review will do; the gate earns its place when prices or stock move daily.

What to take from it:

- The agent answers from your feed, policies and guidelines. Its mistakes are your data's mistakes.
- Check the feed against checkout on every export and keep the last good feed when they disagree.
- Treat the policy text and guidelines as a prompt: write test questions and run them in preview before every change.
- You won't see the conversations, so the checks have to happen before publishing.

If you are wiring a store's catalogue into AI agents and want the feed and the checks built properly, [book a conversation](https://calendly.com/voirlabs/30min).

## Sources

- The Next Web, "OpenAI lets businesses claim ChatGPT profiles and launch their own agents": https://thenextweb.com/news/openai-businesses-in-chatgpt-business-agents
- FoneArena, "OpenAI rolls out Businesses in ChatGPT with Business Hub and Business Agents": https://www.fonearena.com/blog/496812/openai-rolls-businesses-chatgpt.html

---

Voir Labs is an AI-first design and development studio (https://www.voirlabs.io/). Related service: https://www.voirlabs.io/services/ai-automation. Book a 30-minute conversation: https://calendly.com/voirlabs/30min