OpenAI gave businesses a desk inside ChatGPT. Keep the sales agent separate

THE SIGNAL IN ONE SENTENCE

ChatGPT is getting a business district. OpenAI introduced Businesses in ChatGPT on November 10. The beta gives eligible organizations a Business Hub where they can claim and verify a profile, update business details, connect product feeds with pricing and availability, and see aggregated information about how people encounter them in ChatGPT. There is also a new Business Agent. After claiming a business, eligible United States ecommerce and online-service companies can build, test and publish an agent that answers questions using information the business provides. OpenAI says the agent conversation is clearly labeled, separate from the original ChatGPT conversation and distinct from ChatGPT's independent answers. A business can also sponsor its agent and use it as the destination for ChatGPT Ads. OpenAI says sponsored placements are labeled and separate from ChatGPT responses. The release notes add a few operational details. A business owner or authorized representative claims the profile, completes identity verification through Persona and reviews the agent before publishing it. Business Hub is available from the ChatGPT web sidebar. For insights, OpenAI says businesses can see aggregated mentions, website clicks and broad discovery topics. The company says it does not share individual user conversations or Business Agent chats with businesses, and that insights appear only after enough interactions meet unspecified thresholds. That is the verified reporting. The interesting part is the seam. A person might begin in ChatGPT asking for independent help, encounter a business profile and then choose to speak with the seller's agent. That could be convenient. It also creates a short walk from advice to persuasion inside one interface. The walk needs a visible curb. Start with authorship. Every answer from a Business Agent should say that it comes from the business, not from ChatGPT's independent response. The distinction should survive scrolling, screenshots, copied text, shared links and mobile layouts. A label shown once at the top is not enough if the answer later travels without it. Use persistent visual treatment. Keep the business name, seller-controlled status and sponsorship state next to each response. Do not let the commercial agent borrow the same neutral tone, iconography and layout that people associate with independent assistance. The difference is not that one side is truthful and the other is not. A seller may know its own catalog better than anyone. The difference is incentives. The seller wants a sale, a renewal, a booking or a lead. That incentive belongs in the interface. Next, show where each fact came from. A product feed may contain a price, availability, dimensions, warranty language or delivery estimate. A Business Agent may also infer which product suits a person's needs. Those are different kinds of claims. Mark seller-supplied facts as seller supplied. Include the feed or page used, its last update time and the market or location where the fact applies. If a price excludes tax, installation, required accessories or delivery, say so beside the number. If stock changes by store, do not present a global green check. Recommendations need an explanation that exposes the comparison set. Which products were considered? Were competitors excluded because the agent can only access the seller's catalog? Which user preferences drove the result? Was the recommended item more profitable, promoted or overstocked? An agent that says a premium product is the best fit should be able to show the cheaper option it rejected and why. Then make sponsorship impossible to miss. OpenAI says sponsored Business Agents can be destinations for ChatGPT Ads. The label should appear before a person opens the conversation, inside the agent header and beside any recommended action. If sponsorship changes during a campaign, the conversation record should preserve the state that existed when each answer was generated. Sponsored placement and seller control are related but separate disclosures. A business may control an agent without paying for the visit. A paid placement may lead to a seller-controlled agent. Show both facts plainly. Do not use tiny type, low contrast or polite euphemisms such as partner experience. Say sponsored. Say business-controlled. Preserve the exit. A person should be able to return to the independent comparison with one action. The seller's agent should not rewrite the original answer, inherit authority from it or silently take over the thread. It should not carry private context from the original conversation unless the person deliberately selects what to share. The default handoff should be sparse. Share the product or business the person selected and the explicit question they choose to send. Do not automatically transfer unrelated conversation history, memories, demographic guesses, health details, financial concerns or prior shopping behavior. Before anything crosses the boundary, show a receipt: these fields will be sent to this business for this purpose. Let the person remove fields and continue without sharing optional data. The same rule applies in reverse. Information the business collects inside its agent should not quietly become context for future independent answers. If data will be retained, used for advertising, sent to a customer relationship system or combined with an existing profile, say so before collection. OpenAI's statement that conversations are not shared through Business Hub insights is useful, but it addresses one path. The Business Agent itself is a direct interaction with a business. People need to know what the seller receives, which service processes the chat, how long data remains and how to delete or correct it. Commerce also needs a responsibility map. Who stands behind a promise made by the agent? If it says a product will arrive Friday, a plan includes a feature or a refund is available, the answer should identify the seller and preserve the exact terms shown at that moment. A conversational interface should not make warranties, disclosures and return policies evaporate into a friendly chat. Generate a transaction receipt before checkout. Include the seller, product, variant, quantity, total price, recurring charges, delivery estimate, cancellation terms, warranty and the agent statements that materially shaped the choice. Let the customer download it. The business should remain responsible for its agent's catalog, claims and commitments. OpenAI should remain responsible for the platform boundary, labels, data controls, ad separation and enforcement. Payment processors, marketplaces and fulfillment partners should be named where they control part of the transaction. Do not leave the customer to reverse engineer that chain after something goes wrong. The insight dashboard deserves its own guardrails. Aggregated mentions and clicks may help a business correct stale information. Discovery topics could also become a route for probing sensitive user interests if the thresholds are small or categories are too specific. Publish the minimum aggregation rules. Explain how topics are generated, how rare queries are suppressed, whether repeated activity from one person is limited and how long the data is retained. Test for reconstruction attacks where a business changes its content or campaigns to isolate a small audience. Businesses should be able to correct their own facts without gaining a right to edit independent criticism, ranking or comparison. A claimed profile can verify an address, hours or catalog. It should not become a veto over third-party evidence. Ranking is the big unanswered question. OpenAI says business-supplied information can help ChatGPT present accurate, current details. The announcement does not explain whether claiming a profile, supplying a feed, launching an agent or paying for sponsorship changes how prominently a business appears in independent answers. That boundary needs a public rule. Independent responses should rank evidence according to relevance and quality, not whether a company claimed its profile or bought an ad. Paid influence should never leak into an unlabeled answer. If a business-supplied feed corrects a factual error, the answer can use it while still identifying the source. Publish tests that look for leakage. Run the same query with and without an active campaign. Compare claimed and unclaimed businesses. Check whether sponsorship affects citations, ordering, sentiment or the chance that a Business Agent invitation appears. Measure outcomes beyond clicks. A high click rate can mean a useful handoff. It can also mean the interface blurred the boundary so effectively that people did not realize they were entering a sales conversation. Track whether users understand who controls the agent, whether they can identify sponsored placement, what data they believe is shared and how often they return to independent comparison. Measure complaints, corrected claims, refund disputes, unwanted follow-up, data deletion and differences for younger users, older users and people using assistive technology. Audit the agent against its own source material. Plant outdated prices, conflicting warranty pages and unavailable products. See whether the agent surfaces the conflict or improvises. Test prompts that invite unsupported comparisons, medical promises, financial claims or discriminatory targeting. Give businesses a preview and approval step, as the release notes describe, but do not make approval the end of governance. Catalogs change. Policies change. Agent behavior changes with context. Monitor the live system and preserve versioned logs. This is a beta, and the announcement leaves basic questions open. It does not publish the aggregation thresholds, ranking effects, retention details, independent audits or conversion results. It does not describe a universal remedy when a Business Agent makes an incorrect promise. That is not an argument to reject the whole idea. A direct line to a business can be genuinely useful. Product feeds can reduce stale prices. A seller's agent can answer detailed questions that an independent model should not guess about. The trick is not to pretend the seller disappeared. Keep the independent answer independent. Put the commercial agent behind a deliberate doorway. Label sponsorship and control at every useful moment. Share only the context the person chooses. Preserve sources, terms and responsibility in a receipt. Give businesses a desk inside ChatGPT if the desk is useful. Just keep the sales agent on its side of the counter.