Loading

Preparing your content

Article Details
2026-11-17BXT Data Editorial

Promotional feed stalls agent orders

Promotional feed stalls agent orders article image

"Works out of the box" is how OpenAI describes ChatGPT Wallet on existing checkout pages, and the qualifier that follows matters more for e-commerce teams than the launch itself. As of November 17, 2026, the wallet has been live for four days, limited to US ChatGPT Pro subscribers aged 18 and over with an eligible US-issued credit card, and it pays through a network token rather than a card number. What decides whether a purchase completes is duller: a product feed carrying a live price, and a firewall that lets a signed agent through.

What OpenAI shipped on November 13

Known factsUnverifiedTo verify
Launched November 13 for US Pro subscribers on dots; card saved once, biometric approval per purchase, network token replaces card detailsWhether the wallet converts at the same rate as a card-on-file checkoutSession-level conversion for wallet orders, which merchants hold
Commerce Partner Playbook asks for a claimed profile, a product feed with live price and stock, and Web Bot Auth signed requestsHow many retail sites currently block signed agent requests by accidentCDN and firewall logs filtered for ChatGPT agent requests over 30 days
Forwarded header relays asn, for, ip-country, ip-prefix, locale, tz and user-agent to risk systemsWhether fraud providers score agent-initiated orders differently once the header is presentProvider rule sets and decline reasons for tokenised agent orders

ChatGPT Wallet lets an agent pay for a user without holding the card. A shopper saves a card once, reviews the merchant, items, total and selected card before each purchase, and approves with a biometric check; the network then returns a token that stands in for the number, security code and expiry. The merchant receives a network token and can read the payment account reference, the credential covers a single transaction, and the charge lands on the card the shopper chose.

The launch is narrow on purpose. It runs through dots for US Pro subscribers aged 18 and over with an eligible US-issued credit card, built with American Express, Chase, Mastercard and Visa, and OpenAI states no fee for accepting a wallet payment while leaving pricing, fulfilment, returns and support with the merchant. Narrow scope does not reduce the preparation work, because that work sits in product data and risk rules rather than in a new integration.

The feed is where price consistency is decided

One item, one identifier, one price

The Commerce Partner Playbook asks a business to claim its profile in ChatGPT, verify it, then upload a product feed carrying current price and availability; the feed can arrive by manual upload, hosted URL or SFTP, and it does not update a cart or confirm an order. That last clause is the important one: the feed is a discovery surface, so a stale price there is a promise the checkout will contradict, and the resulting price gap appears at the worst possible moment.

Availability is part of the price

An agent that reads a feed where a variant is listed as available will attempt the order. If the variant is gone, the shopper sees either a substitution or a failure, and both are recorded against the merchant. Keeping availability honest at SKU level is the less obvious half of price monitoring, because a discount on an item that cannot ship reads as bait rather than as a promotion.

FieldWhat the agent needsWhat breaks without it
Item identifierOne stable ID across feed and checkoutDuplicate listings and split review history
PriceLive price including any promotionFeed price contradicts the cart
AvailabilitySKU-level stock by locationSubstitution or a failed order
Variant attributesSize, pack and colour in plain textWrong item added to the cart
Delivery termsStructured cost and windowSession ends at the shipping step

The merchant keeps the customer relationship, so the usual levers stay in place: pricing, promotion depth, returns and service. What changes is who reads the page first. If the agent reads a promotional price the checkout cannot honour, the discount has already been advertised by the time the decline happens, and that session never appears in the merchant's funnel report.

Firewalls, signed requests and the decline problem

The second failure mode is access. OpenAI signs certain browser requests with Web Bot Auth under RFC 9421, sending Signature, Signature-Input and Signature-Agent headers set to chatgpt.com, and publishes keys so a CDN or firewall can validate them; Akamai, Cloudflare, HUMAN and Vercel verify automatically, and other providers are told to check the signature before adding an exception. Teams that tightened bot rules against scrapers are the most likely to be blocking a paying agent by accident.

The third is risk scoring. A purchase through a cloud browser can trip fraud models tuned on IP address and device signals, so OpenAI is rolling out a Forwarded header that relays seven parameters, among them asn, ip-country, ip-prefix, locale, tz and user-agent, and names Forter, Riskified, Sardine and Signifyd on the fraud side with Adyen, Checkout.com and Stripe among processors. A declined agent order still registers as the merchant's failure.

A four-step readiness pass

  1. Export the category feed and confirm every item carries one stable identifier used on its product page.
  2. Compare feed price and stock against the checkout for the ten highest-traffic SKUs.
  3. Pull CDN and firewall logs for blocked ChatGPT agent requests over the last 30 days.
  4. Ask the fraud provider how a tokenised, agent-initiated order is scored, and write the answer down.
FieldExampleWhere to check
Feed item IDSKU-4471-BLKFeed export against the product page URL
Feed price$38.00Compare with the cart at 09:00 local time
Stock flag1 of 3 variants in stockAvailability API
Blocked agent requests412 in 30 daysCDN log filter by signature header
Risk decision on agent orderManual review, not declineProvider rule set

The pass takes an afternoon per category and produces a list of defects rather than opinions. A feed whose identifiers change when a variant is added will keep splitting sessions, and a firewall rule that returns a challenge page to signed requests will keep looking like a model failure in the agent logs.

Sequence matters more than tooling. Fix identifiers and price first, because no amount of access work rescues an order built on a stale price; confirm agent access second; align risk scoring third. Doing it in the other order produces traffic that arrives and fails, which is the most expensive version of the problem.

FAQ

  • Does ChatGPT Wallet replace the merchant's checkout?

    No. It works with existing checkout pages and standard card forms, and the merchant receives a network token rather than a card number. No wallet-specific integration is required, which is why the preparation work concentrates on feed quality and risk rules.

  • What is the Forwarded header for?

    It relays signals about the real shopper, including network, country, locale, time zone and browser, so a fraud model can score an agent-initiated order on evidence about the person instead of on the cloud browser's address.

  • Where does a price gap show up first?

    In the feed, because the agent reads price before a cart exists. If the feed carries a promotional price the checkout cannot honour, the gap is created at discovery and surfaces only at payment, after the shopper has already committed attention.

Sources