OpenAI Failed at the Checkout. Now It Wants the Conversation Before It
OpenAI couldn't take a cut of the sale inside ChatGPT, so it's settling for something sneakier: owning the conversation where the sale gets decided. Business Hub — verified profiles, business agents, analytics — launched Tuesday, and it's less a product launch than a retreat dressed up as a strategy.
The checkout died. The pitch moved upstream.
In September 2025, OpenAI launched Instant Checkout with Stripe and the Agentic Commerce Protocol, promising a commission on every sale made inside the chat. Six months later it quietly killed the first version, admitting it "did not offer the level of flexibility" merchants wanted. The transaction went back to the merchant's site.
What OpenAI kept was the half nobody was watching: discovery. The browsing, comparing, and deciding that happens before anyone pays. Business Hub is built entirely on that half.
What businesses actually get
- A verified profile with a blue check — claim it, prove you're authorised, correct your details. Retailers can add self-service product feeds with pricing and availability.
- A Business Agent that answers customer questions inside ChatGPT — but only for eligible ecommerce and online-service businesses in the US, and only when the customer chooses to start the conversation from the profile.
- A dashboard of aggregated mentions, impressions, website clicks, and "discovery topics" over the past 28 days. What you don't get: the conversations themselves. OpenAI shows aggregated figures, not individual chats.
Listen to the words OpenAI chose
"Discover, compare, decide." Not buy. And the person making the case to journalists was Benji Shomair, VP of Monetization — which suggests, without proving, that this is a revenue strategy wearing a product feature's clothes.
The model is now obvious: the sale happens elsewhere, the decision happens in ChatGPT, and the business gets a verified voice in it plus a count of how often it comes up. That's the mechanism. It is not, yet, a business model.
The honest objections
- The tools are free — for now. OpenAI has said nothing about what businesses will eventually pay for: profiles, agents, analytics, or placement in answers. Sponsored Agents already exist in the product, so paid business conversations are already here. The price tag is just deferred.
- A blue check buys no placement. OpenAI says claiming a profile doesn't guarantee higher placement or more sales, and ChatGPT keeps making its own independent recommendations. That's a promise, not a mechanism — and the verification process deciding who gets the check remains largely undisclosed.
- This might just be feature parity. Google's business profiles work because Google is where search happens. ChatGPT's discovery volume is unproven — 1.2 billion weekly users is a headcount, not a count of shoppers. Whether people actually discover products in ChatGPT in numbers that make these profiles valuable is the open question the whole bet rests on.
The checkout, in its first form, is gone. The conversation is the product now. What it's worth is the next thing to find out.
Go deeper
Read the full analysis on Agentic Affairs: OpenAI Failed at the Checkout. Now It Wants the Conversation Before It
Originally published on Agentic Affairs (https://agentic-affairs.com)
Top comments (6)
Owning checkout meant taking on merchant-of-record friction, refund disputes, and rigid catalog schemas across sellers who already have Stripe or Shopify wired up. Moving upstream to discovery avoids all of that operational liability while keeping the highest-margin layer.
The 28-day aggregated dashboard is the revealing piece. Because OpenAI withholds individual conversation logs, merchants can see impressions and outbound clicks, yet they have no clean way to separate organic intent from what would have converted anyway. Once Sponsored Agents expand, sellers end up bidding for visibility inside a funnel where the platform controls both the recommendation and the attribution window, much like early search-ad directories before third-party tracking matured.
The attribution gap is the part that should worry merchants most — they can see the clicks, but without the conversations they can't tell whether ChatGPT created the intent or just intercepted a sale that was happening anyway. And you're right about the shape of what's coming: once Sponsored Agents expand, sellers are bidding inside a funnel where OpenAI controls both the recommendation and the window it gets measured in. Do you think merchants push back and demand third-party measurement before they pay, or does the traffic volume make them accept platform-reported numbers like early search did?
Most merchants accept platform-reported numbers early on because ad auctions punish holdouts. If the first cohort of sellers sees cheap attributed conversions, competitors bid to avoid ceding the recommendation slot even when incrementality is unproven. Third-party holdout audits only gain leverage later, once auction clearing prices push marginal customer acquisition cost close to gross margin.
That's the mechanism I was missing — the auction itself enforces acceptance, since holding out means ceding the slot to a competitor who bids. Do you think the leverage flip is measurable in advance, like attributed CAC crossing some share of gross margin, or is it only obvious in hindsight once prices have already risen?
The hidden friction in optimistic routing is that failure carries an asymmetric context penalty. When an agent tries the ten-cent model first and hits a tool-call retry or silent state degradation, the conversation history inflates before the session escalates to the eight-dollar tier. Escalating with four times the accumulated context tokens to an eighty-times more expensive model rapidly wipes out the savings from the successful cheap runs. Unless the router can discard tainted turns and re-anchor from clean state, optimistic tiering turns into an expensive re-run tax.
The re-run tax is a sharp way to put it — the cheap run only looks cheap until the escalated retry carries four times the context at eighty times the price. This piece doesn't get into routing design, so I won't pretend it answers that, but the discard-and-re-anchor idea is the crux: does any router you know actually do that in practice, or is the tainted context mostly carried forward as-is?