AI Knowledge Hub
Menu
Back to the libraryEssay / Commerce & growth

Found. Chosen. Paid. The Marketplace Emerging Inside ChatGPT

Business Hub, ChatGPT Ads and Wallet are taking shape as a marketplace. A missed delivery promise exposes the gaps between being found, chosen and paid.

It’s December 12. Suppose your bestselling coffee grinder, the compact one people buy as a gift, just lost its guaranteed Christmas delivery date. Your team fixes the product page in an hour.

But your grinder no longer lives only on your product page. Its price and availability can appear in a product feed. Two separately configured agents can answer questions about delivery. An ad can send shoppers to the old offer. And an eligible shopper can approve a purchase in ChatGPT while an agent completes checkout on your site.

How many of those places learned that the promise had changed? And which decisions should stop while the old promise is still circulating?

The two handbooks

OpenAI says the Sponsored Business Agent’s settings currently do not sync with the agent in Business Hub. You have two colleagues answering the same delivery question from two versions of the handbook. The Business Hub agent can explain an offer and link to the merchant’s site. Use both builders to check which delivery answer each agent gives after a correction. Sponsored Business Agent setup · Business Hub agent capabilities

An accurate answer may end the sale. When the grinder will arrive too late, the agent that tells the shopper to look elsewhere has done its job.

What OpenAI controls along the sale

In November, Business Hub joined ChatGPT Ads and Wallet arrived in dots. Put them beside one another and a marketplace starts to take shape. A merchant supplies an offer, an auction sells a chance to be chosen, and an agent can carry out the approved payment.

InterfaceOpenAI’s productWhat it controls
FoundBusiness HubHow a merchant can supply business and product information for discovery
ChosenChatGPT AdsWhich eligible offers get a sponsored placement while a shopper is deciding
PaidWalletHow a shopper authorises an agent to pay at the merchant’s checkout

Found. Chosen. Paid. It is an appealing sequence. Then you look at who can use each part.

Wallet’s initial rollout is limited to eligible US adults on Pro, in dots, with supported credit cards. Pro is ad-free. Sponsored placements reach eligible Free and Go users, who can buy on the merchant’s site. The account eligible for the new Wallet payment route sits outside that ad audience. Wallet availability · ChatGPT Ads eligibility

That is the first break in the picture. The products line up as a marketplace, but their current audience rules interrupt the journey from being chosen through an ad to paying through Wallet.

The architecture is almost boringly classical. Amazon has supply, demand and payment. Google has them. The fight is over where the intent forms. In ChatGPT, a shopper can work out what they want in natural language, across a conversation, while the assistant helps turn a vague gift idea into a purchase decision. That is the moat OpenAI is trying to build. The assistant can become the place where the market takes shape. OpenAI’s ad-stack engineering account

The part everyone forgets: this is the second attempt

In September 2025, OpenAI launched Instant Checkout with single-item purchases inside ChatGPT. In March 2026, Walmart’s Daniel Danker told WIRED that the selection sold directly inside ChatGPT converted at about one-third the rate of the selection that required a click out. He pointed to fragmented single-item checkout as a problem for shoppers with an existing Walmart cart. The product selections differed, so the comparison leaves open how much of that gap came from checkout. Instant Checkout launch · WIRED’s interview with Danker

OpenAI gave its own explanation for the change in March. Its initial checkout design had lacked flexibility, and merchants would be able to use their own checkout experiences. With Wallet, an agent drives your existing checkout with a network-issued token. No Wallet-specific integration required, though site access and payment compatibility still matter. The “works out of the box” claim carries a lesson from that earlier attempt. The merchant already has a checkout that joins the basket, payment and order together. Wallet asks the agent to work through it. OpenAI’s March update · Merchant payment guidance

Remember this when someone tells you agentic commerce is inevitable because OpenAI shipped a wallet. The new architecture is another bet about how people will buy. Whether it converts better is still an open question.

Late, but industrialising

The race was already under way. Google announced its Business Agent and Universal Commerce Protocol in January. Meta introduced its Business Agent in June. Amazon opened its Shop Direct Merchant Workspace in November, extending a service it had launched the previous year. Google has also announced native checkout for eligible merchants using sponsored Direct Offers. Google’s commerce announcement · Google’s Direct Offers expansion · Meta’s Business Agent · Amazon’s March update · Amazon’s merchant workspace

But watch the velocity. OpenAI reported a $1 billion annualised ad revenue run rate in under 200 days at the end of August. By October, it reported more than 100,000 advertisers had run campaigns and ads were available in 89 countries. That is a substantial ad business to bring into this race. The gap between its scale and Wallet’s narrow rollout makes the audience split more striking. OpenAI’s August milestone · OpenAI’s October expansion

Inside the machine

The model nominates. The auction ranks and prices.

The design in OpenAI’s engineering post should change how you think about the ad system. A post-trained frontier LLM runs an agentic loop of tool calls. It tracks your intent across conversation turns, rewrites the conversation into retrieval queries, generates embeddings, runs approximate-nearest-neighbour search over multiple specialised indices, filters candidates for eligibility, budget and pacing, scores them for relevance and predicted engagement, and nominates them to a separate auction tool that handles admission thresholds, quality-adjusted ranking and pricing. Ad-stack engineering account

Translation for operators:

  1. Retrieval is semantic. The system embeds rewritten intent queries to find candidates in specialised indexes. Accurate product data, clear ad copy and useful context hints help describe when an offer fits. Feed engineering starts to look like part of the discovery work marketers already do for search. Targeting and context hints
  2. Relevance gates admission. The auction applies a minimum threshold for relevance and value. A larger bid still has to meet that bar.
  3. No ad can be a successful outcome. The system can decide that a conversation has no suitable placement. An ad can also arrive too late or fail an eligibility check. A quiet campaign leaves a question to investigate before raising the bid. Serving diagnostics
  4. Codex helped engineers build the stack. OpenAI describes guardrails for shipping code and a separate evaluation loop in which stronger models assess delivery and provide training feedback to other ads models. Human and advertiser feedback also shapes the changes. The system is learning on several levels at once. Engineering workflows and feedback

Where the money flows

The wallet’s design is quite conservative:

  • You save an eligible card once. A payment service provider collects the card details. OpenAI receives a tokenised credential and limited card information, including the last four digits. It does not receive the full card number or security code. Wallet Privacy Notice
  • At approval, the payment network issues a substitute token, with a card number, security code and expiry, for the agent to use at checkout. The merchant and agent receive these substitute credentials. Wallet announcement
  • The agent completes the merchant’s own checkout. The merchant keeps pricing, fulfilment, returns, fraud decisioning and customer service. Merchant payment responsibilities

The agent interprets, the user authorises, the merchant fulfils, the network supplies payment credentials, and OpenAI operates the surface. Follow a purchase across those responsibilities and another handoff appears. Someone still has to confirm that the order exists.

What can be measured across the handoff?

The next connection to look for is the feedback loop. It is easy to picture one running all the way through. A shopper buys, the platform records the outcome, and the bidder learns which opportunities are worth paying for. Wallet does collect transaction and order information to operate the service. Its Privacy Notice says that information is neither sold nor shared for cross-context behavioural or targeted advertising. Meanwhile, OpenAI documents merchant-reported conversion events for ads. The public guidance gives us no documented Wallet-to-bidding integration. A purchase record and an ad signal have different jobs, even inside the same company. Wallet Privacy Notice · Conversions API

Even the word “conversion” changes with the report. General Ads API Insights can include outcomes attributed to clicks and views under the selected windows. The Conversions API guide describes Ads Manager’s main conversion total as click-through, with view-through conversions reported separately where available. Two colleagues can arrive at the same meeting with different conversion totals and both have copied their numbers correctly. The report, event and attribution window tell us what they actually counted. Ads API reporting · Conversions API reporting

What to check on Monday

The arithmetic of a $25 Target CPA

Say your grinder clears $35 in contribution per additional retained order after variable costs and returns. You set a $25 Target CPA and observe an actual cost of exactly $25 per credited purchase.

At that rate, 100 credited purchases cost $2,500 in media. Divide by $35 and you need at least 72 additional retained orders per 100 credited purchases to cover media, before overhead. Some credited purchases would have happened anyway. Some will be cancelled or returned. If enough of those orders fall away, the campaign can meet its target while losing money. The grinder bought on an expired delivery promise can count towards a successful CPA long before the cancellation reaches your books.

Use your recent actual CPA for the same event and a reporting window that allows for conversion delay. If you send a purchase through both the browser Pixel and Conversions API, use the same Pixel ID and match the browser’s event_id to the server’s id. Diagnose missing or duplicate events before drawing conclusions from the new strategy. OpenAI says weak signals can restrict delivery. Better matching can also raise the number of purchases credited to ads without changing how many grinders you sold. Use merchant order records to check retention and a credible comparison to estimate additional sales. Bidding guidance · Conversion event deduplication · Christmas pilot and lift test

What the campaign may change

Review what your campaign permits OpenAI to change, too. Creative enhancements are on by default for self-serve and BPO campaigns from October 30, subject to existing opt-outs and campaign overrides. Enabled enhancements can adapt copy and destinations. Check the effective settings where you run the campaign and make sure each eligible destination still describes an offer you can fulfil. Creative enhancements and controls

The point where approval becomes an order

The Wallet help guide tells shoppers to check for merchant confirmation before trying again if a result is unclear. That’s the handoff to rehearse. A shopper can approve the purchase, the checkout can time out, and the merchant may already have an order. Your support team needs a way to check before asking the shopper to retry. OpenAI also documents how to recognise signed browser requests at your edge and notes that a payment token can look different from the shopper’s physical card number. Check agent access, risk handling and order lookup with your own providers. Wallet purchase and support guide · Signed browser requests · Payment and risk handling

The seams: five gaps you can work

By now, the gaps have become part of the marketplace’s shape. The merchant can supply facts without knowing how much they influenced an answer, buy attention without reaching Wallet’s launch audience, and receive a purchase event before knowing what the order will contribute. Five seams are worth following.

  1. Organic feed influence. A corrected feed gives ChatGPT better facts to draw on. It can still use other sources, and submission carries no recommendation guarantee. The merchant can improve what it supplies while the effect on discovery remains hard to isolate. The grinder’s new delivery date gives us one specific correction to look for. Business Hub setup and discovery
  2. The measurement gap. The conversion total can improve while retained sales stand still. Server-side events help capture and match orders. Holdouts can estimate how many sales advertising added. Marketing mix modelling can examine channel contributions where the data supports it. Those methods answer different questions, and none establishes a Wallet-to-bidding connection. The opportunity is to find out which improvement the business actually got. Conversions API · Christmas pilot and lift test
  3. The audience split. The Pro shopper with Wallet sits outside the current ad audience. Organic discovery and business information become especially relevant to that route. The merchant selling the same grinder has two ways a shopper can arrive, with different evidence of what brought them there. Wallet availability · ChatGPT Ads eligibility
  4. Verification and upkeep. Shopify offers a streamlined connection, but Persona approval still gates changes, feeds and Insights. A merchant can have a verified domain and still be waiting to manage its information. Even after approval, somebody has to keep the product facts and the two handbooks current. Free tools still create work. Business Hub verification
  5. Insights blind spots. For an online business, a website reference counts as a mention. A brand name alone can leave no recorded mention. A blank metric can mean activity is below the threshold, collection is ongoing or reporting is unavailable. Local results follow a different counting rule. The merchant checking whether the corrected offer reached shoppers gets a view with limits of its own. How Business Hub counts activity

These gaps already shape what merchants can do, and what they can verify.

OpenAI has described a future in which shoppers pre-authorise purchases within limits such as budget, category and time. If that delegation expands, an outdated promise could travel further before a person catches it. Today, each Wallet purchase still requires approval. How much uncertainty can the system resolve before asking the shopper to act? Wallet roadmap · Current purchase steps

The deadline belongs to the promise

There are two published dates for the bidding transition.

  • November 18, 2026. Under OpenAI’s published transition, using the legacy bid fields for API campaign creation or an ad-group bid-type change selects Target CPA or Target CPC from this date. The submitted value becomes an average-cost target for those operations. Check the resulting strategy and value before scaling any automated changes. API transition guidance
  • January 1, 2027. OpenAI schedules Max Bid and Max CPC retirement for existing campaigns and deprecation of the legacy API fields. Bidding migration guidance

For the grinder, Christmas supplies a different deadline. Once the delivery promise expires, every old copy can sell an offer the warehouse can no longer fulfil. That is the reason to rehearse the grinder test before your own cutoff. Change one delivery promise on a controlled offer and record when the product page, feed, agent answers, ads and checkout reflect it where you can observe them. If an ad is still making the old promise, pause the affected campaign while you correct it. At checkout, verify the current delivery terms and ask the shopper to review any material change before the order proceeds. Then check where your team would look if a customer said they approved payment and never got a receipt.

The grinder brings us back to a disagreement about success. The agent can give an accurate answer that ends the sale. The bidder can hit its target on an order that later disappears. Wallet can record approval while the shopper is still waiting for confirmation. Each has completed a different part of the job.

The merchant lives with the whole result. The grinder test is a way to find out whether anyone sees the missed promise before the customer does.

OpenAI’s product and API documentation for this article was checked on November 19, 2026. The grinder scenario and its figures are illustrative.