ChatGPT Wallet does not give OpenAI a checkout of its own. A dot fills in your existing checkout page and pays with a network token, while your business keeps the order, the fulfillment, the support and the fraud decision. Whether you get these orders comes down to two systems you already run: the bot manager in front of your site, and the fraud rules behind your payment form. Both were tuned for a world where an automated browser holding a card meant trouble.
OpenAI launched ChatGPT Wallet on Friday, November 13, to a limited number of US ChatGPT Pro subscribers who are 18 or older and hold a US-issued Visa, Mastercard or American Express card. Volume will be tiny for now. The configuration work is small too, which makes this the right month to do it.
How Does ChatGPT Wallet Pay a Merchant?
ChatGPT Wallet pays with a payment token that stands in for the card, so neither the dot nor the merchant sees the card number. According to TechCrunch's launch report, the user adds cards once, asks a dot (OpenAI's name for its agents) to buy something, reviews the items, total and card, and approves with Face ID or a fingerprint. The dot passes the merchant what it needs to ship the order, such as the address and contact details. Card rewards still accrue and standard dispute rights still apply.
The merchant side is simpler than most agentic commerce pitches. OpenAI's Accept payments guide says Wallet "works out of the box with existing checkout pages and payment forms" and that "no Wallet-specific integration is required." The customer approves in ChatGPT and your checkout processes the payment. In OpenAI's words, "Your business handles the order, fulfillment, customer support, and fraud decisions."
A network token is a card-network substitute for the card number, issued for a specific use, that your processor can authorize like a card. OpenAI tells merchants to use the Payment Account Reference (PAR) to tie that token back to the underlying account. PAR is an EMVCo data element that the PCI Security Standards Council describes as a value that lets "acquirers and merchants to link tokenized transactions" to transactions on the same card. PCI does not treat PAR on its own as account data, and it cannot start a payment.
Three rules in the guide will catch out a checkout built for humans:
- The credential is single use. "The customer consents only to the specific transaction," and OpenAI says "Don't reuse or save that payment credential for a separate purchase." If your checkout auto-saves the card to the account on every order, that default now breaks OpenAI's terms on Wallet orders.
- Subscriptions need their own consent. Recurring billing is allowed if your checkout already supports it, with consent captured and your processor's merchant-initiated transaction practices followed.
- Existing card-on-file flows continue. Wallet is a separate payment option, so returning customers with saved cards are unaffected.
The product is also incomplete. TechCrunch reports that spending limits, category and time-window controls and pre-approved purchases are planned but not yet live. OpenAI's plugin guide says "We're still working on the interaction patterns between third-party plugins and ChatGPT Wallet." Today, a dot shopping your site is a cloud browser on your normal checkout.
Will Your Bot Manager Block ChatGPT's Agent?
Probably not by default, but your own custom rules might, and OpenAI documents the exact setting to check for each major vendor. The browser access guide says ChatGPT's cloud browser "can use your existing purchase or booking flow" and points merchants to change settings only if their security controls block verified visits.
The identity mechanism is Web Bot Auth: requests are signed using HTTP Message Signatures, the IETF standard published as RFC 9421 in February 2024. Per OpenAI's commerce documentation, signed requests carry Signature, Signature-Input and Signature-Agent: "https://chatgpt.com", and the public keys sit at https://chatgpt.com/.well-known/http-message-signatures-directory. If you verify manually, OpenAI says to confirm the Signature-Agent value matches exactly, including the quotes, and to verify the signature before applying any allowlist exception. A user agent string alone proves nothing.
The vendor settings, from the same documentation:
| Bot manager | Where the setting lives | What to set |
|---|---|---|
| Akamai | Bot & Agent Control | Custom bot category with the Verified Bots entry "ChatGPT Agent", set to Allow |
| Cloudflare | AI Crawler and Agent access policies, or a WAF custom rule | "dots, OpenAI" under Personal Agents set to Allow, or detection ID 129220581 (bot tag chatgpt-agent) |
| HUMAN | Sightline, Known bots & Crawlers | "ChatGPT Agent" ON and Allow |
| Vercel | No change unless your rules block it | Identified as chatgpt-operator |
OpenAI's Agent Access page says Akamai, Cloudflare, HUMAN and Vercel each verify the signatures automatically, so a team on one of them does not need to write verification code.
This is separate from crawler policy. OpenAI's crawler documentation lists OAI-SearchBot and GPTBot with their own robots.txt controls, and says robots.txt rules may not apply to the user-triggered ChatGPT-User. A retailer that blocked GPTBot to keep content out of training has made no decision about Wallet, and should not reuse that rule for the agent.
The case for caution is real. Cloudflare's signed agents documentation now labels services like this "Intermediary": many end users operate them, and "each action is initiated by a different end user." Trusting the signature tells you the request came through ChatGPT. It does not tell you the person behind it is a good customer. That is why the allow setting should open your checkout to the agent without exempting it from fraud scoring.
What Changes in Your Fraud Rules?
Your fraud stack keeps the decision, and it needs new signals to make it, because the browser, IP and device belong to OpenAI's cloud and not the shopper. OpenAI says it is relaying user signals in the Forwarded HTTP header: asn, for, ip-country, ip-prefix, locale, tz and user-agent, per its Accept payments guide. It tells merchants on a third-party risk provider to confirm the provider ingests those fields, and merchants with in-house systems to start observing the header. Cloudflare is experimenting with the same header (RFC 7239) so site owners can apply rules to the party behind an intermediary request.
Without those fields, a velocity rule sees hundreds of different cards arriving from a handful of datacentre addresses with identical browser fingerprints. To a rule set written for human shoppers, that looks like card testing.
Disputes are the open question. OpenAI's payments guide does not address chargebacks. Chargeflow, a chargeback-management vendor with an interest in the problem, argues that an agent-placed order is disputed under the same fraud and "not as described" codes as any other, and that "Visa has not published how disputes are treated" under its agent protocol. Treat that as one vendor's reading of the rulebooks and confirm it with your acquirer. Either way, your representment evidence has to show what the customer approved: keep the PAR, the signed-request headers and the order details the dot submitted, linked to the order ID.
Identity is a separate track. OpenAI's commerce docs say Sign in with ChatGPT is limited to select partners and warn that "a ChatGPT identity is not purchase approval." Do not let a ChatGPT login stand in for the payment approval step in your own logic.
Should You Submit a Product Feed Now?
Yes, if ChatGPT is already sending you traffic, because the feed is how a dot finds your products and you control what it says. Per OpenAI's product feed guide, you start in Business Hub, verify your identity (required before upload), pick your countries and currencies, and deliver by manual upload, a hosted HTTPS URL or SFTP. The guide is explicit that "the feed does not update a cart or confirm a purchase"; your live site still confirms the price and stock at checkout. OpenAI's documentation adds that setting is_eligible_checkout alone does not complete checkout onboarding, which needs a separately enabled integration.
Measurement is the gap. Business Hub, which we covered when it launched, counts clicks and not purchases. Wallet purchases will show up in your own order system, and you will only be able to attribute them if you tag the agent traffic yourself.
The strongest argument for waiting is the demand data. TechCrunch notes that DoorDash and Uber have reported low numbers of agent-placed orders, a gap we measured against Uber's 320 million weekly trips. Amazon, the largest retailer that bans AI agents, steers shoppers to its own assistant instead, and has blocked agents before. Agents also still make mistakes: Visa's own benchmark found models completing purchases after failing the task. A cautious retailer can reasonably decide the volume does not justify the risk yet.
The counter is cost. Checking one bot manager setting and adding seven header fields to a fraud model takes days, and you get the first data while order counts are small enough to review by hand.
The Payment Rails Are Converging on Signed Agents
ChatGPT Wallet puts a card network's token and an RFC 9421 signature into one ordinary checkout. Visa's Trusted Agent Protocol, announced October 14, 2025 and built with Cloudflare, rests on the same HTTP Message Signature standard and lists Adyen, Checkout.com, Shopify, Stripe and Worldpay among its early partners. Meta's competing approach, which we looked at in October, runs through OAuth instead, and banks such as Schwab and JPMorgan have blocked agents that use customer logins.
The pattern resembles the early years of card-not-present ecommerce, when merchants learned to accept orders from customers they could not see by building fraud models on new signals. The signals this time are a signature, a token and a forwarded header, and OpenAI has documented all three.
This Week:
- Ask your bot management owner which row of the table above applies to you, and pull a week of logs for
Signature-Agent: "https://chatgpt.com"to see whether your rules already challenge or block it. - Ask your fraud provider in writing whether it ingests the
Forwardedheader fields OpenAI lists, and when.
This Month:
- Have payments engineering confirm your checkout does not save a Wallet network token to the customer account, and that PAR is stored with the order.
- Write a fraud rule that scores signed ChatGPT sessions on the forwarded signals instead of the datacentre IP, and route the first few weeks of those orders to manual review.
- Claim your Business Hub profile and submit a feed if ChatGPT already appears in your referral reports. If your enterprise also runs OpenAI's Dots for employees, keep that admin decision separate from this one.
This Quarter:
- Agree with your acquirer what evidence wins a dispute on an agent-placed order, and make sure your order system keeps it.
OpenAI says spending limits and pre-approved purchases are coming next, and those will let a dot buy with less review by the customer each time. Have the allow rule and the fraud rule in production before that ships.
Continue Reading
- OpenAI's Business Hub Counts Clicks, Not ChatGPT Purchases
- Your AI Shopping Agent May Buy After It Fails, Visa Finds
- Third-Party Agents Brought Uber Four Rides in a 320M-Trip Week
- Amazon Blocks Meta's Muse Agent Where the Perplexity Ruling Stops
- Meta's Personal Agent Protocol Signs Walmart Before It Has a Spec
- OpenAI Starts Billing for ChatGPT Enterprise Dots on November 18
