ChatGPT now lets brands claim a profile and run their own agent. Would you put yours there?

by•

For a lot of products, the first real conversation a customer has might soon be with your agent inside someone else's app. I'm not sure how I feel about that, and I'd like to hear how other makers are thinking about it.

This week OpenAI opened a beta it calls Businesses in ChatGPT. From OpenAI's announcement as summarised by and Dynamic Business (OpenAI's own page wouldn't load for me, so this is secondhand), it has two parts.

Business Hub lets a verified business claim its ChatGPT profile, fix outdated information and connect a product feed with prices and availability. It's open in markets where ChatGPT is available, and verification is required before you can publish changes, connect a feed, see analytics or build an agent. Shopify stores can connect through the Shopify admin.

Business Agents are the more interesting part. Eligible US ecommerce and online-service businesses can build an agent that answers questions about their products and policies, inside limits and a knowledge base the business sets. A shopper comparing options in ChatGPT can step into a brand's agent, which opens as a separate, labelled conversation, so the independent answer and the company's own representative aren't mixed together. OpenAI says it doesn't share individual conversations with businesses, including agent chats; the Hub shows aggregate numbers like mentions, profile views and referral clicks once activity passes a privacy threshold. Businesses can also sponsor their agent as a destination for ChatGPT Ads, with labelled placements, and the reporting says claiming a profile doesn't buy better treatment in ChatGPT's normal answers. Reports describe it as free during the beta, with paid tiers not ruled out.

I don't want to argue about whether this is good or bad for the web. What I keep turning over is a smaller, practical question: what changes for a small team when the first conversation about your product might happen somewhere you don't control?

A few things I keep coming back to.

Your landing page stops being your first impression. If someone has already read ChatGPT's summary of you and asked your agent three questions, they arrive at your site with a formed opinion. The page has to confirm something, not introduce it.

The agent's limits become your brand. A human support person can say "I'm not sure, let me check." An agent built from a knowledge base will answer whatever it's asked unless you deliberately tell it not to. What you forbid it to say is probably more important than what you teach it.

You lose the transcript. That's the right call for users' privacy, and I'd want it that way as a user. But it means the single best source of product feedback, the exact words people use when they're confused, goes into a dashboard of topics instead of your inbox.

For Murror this is harder than for most. People come to us to understand how they feel, and an honest answer to "will this help me?" is "maybe, and here's what it isn't." I'd want an agent that's happy to say "this isn't the right tool for that" and point somewhere better. I'm not sure a sales-shaped surface rewards that, and I'd rather not ship an agent than ship one that oversells.

So, open questions, and I'd love people to answer each other, not just me:

Would you claim your profile and build an agent, or wait and see? Why?

If you did build one, what's the one thing you'd never let it say on your behalf?

And for anyone who already runs a support bot on their own site: what did you learn about where it should hand over to a human?

73 views

Add a comment

Replies

Best

From a small seller's side I would claim the profile first and wait on the agent. Claiming and correcting the basics (price, availability, what is actually in the box) is cheap, and a wrong fact repeated in someone else's app is hard to fix later. For the agent, the one thing I would never let it say is a promise the team cannot back: a delivery date, a stock level, a refund exception. A small shop cannot honour what a bot invents on its behalf. On handover, I would start strict: anything about an order that already exists, a complaint, or a custom request goes to a human, and the agent says so plainly instead of guessing.

 claiming first and waiting on the agent is where I've landed too. Fixing a wrong fact costs almost nothing now and a lot later. I like your rule about promises the team can't back, because it gives you a simple test for any answer: could a person on the team actually honour this tomorrow? And "anything about an order that already exists goes to a human" is a cleaner handover line than most I've seen. Do you think the agent should also say why it's handing over, or does that just add noise?

  Yes, one short line is worth it. Saying why it is handing over ("this is about an existing order") shows the customer it is not a dodge, and it lets the person on the team skip the re-asking.

 that second benefit is the one I hadn't thought about. The reason line helps the customer, and it also works as the handover note for whoever picks it up, so nobody has to ask the same question twice. One short line it is. Thanks for coming back to answer.

we build website agents at Tobira, so your transcript point hit home. reading our agent's chats is where we found the real miss: answers were right, but it never offered a next step, so people just left. a dashboard of topics wouldn't have shown that. the one thing ours never says is a price that isn't public, anything custom goes to a call with a person

 "answers were right, but it never offered a next step" is a great example of something only a transcript shows you. Every answer would have passed a correctness check and people still left. It makes me think the thing to ask for from a surface you don't control is some sample of anonymised full conversations, because a topic count can't show you that kind of miss. How often do you still read chats line by line now that the agent has been running a while?