How the audit works
We load the homepage and up to five product pages once each, in a fresh headless Chromium browser with its normal user agent. We never click, type, submit forms, create carts, open checkout, or imitate ChatGPT's signed requests. We respect robots.txt for our own scanner and stop if a site asks for less traffic (HTTP 429).
Statuses. Pass and fail mean we observed it. Warning means an issue was observed but is not a confirmed blocker, or only some sampled pages passed. Unknown, not tested and not applicable are excluded from the score. "Needs merchant verification" items are never counted either way.
Score. Four groups — access (20), products (35), cart pathway (25), policies and trust (20). Pass earns the check's points, warning half, fail zero. Score = earned ÷ evaluated points. We show no score when coverage is under 70%, fewer than two product pages could be read, or the scan was blocked.
What a public scan can't prove. Whether ChatGPT's verified browser is allowed through your firewall, whether your payment processor accepts a Wallet token, whether fraud rules cancel agent orders, or whether a product feed is submitted. Those need the merchant.
Not affiliated with OpenAI. Checks follow OpenAI's published Commerce Partner Playbook as of the review date above.
| ID | Check | Weight | |
|---|---|---|---|
| A01H | Homepage reachable An agent starts where a shopper starts. If the homepage errors or times out, nothing else can happen. | Accessible site and browser-readable pages · 5 pts | Source |
| A01P | Product pages reachable An agent has to open the product page to read price, options and stock before it can add anything to a cart. | Accessible site and browser-readable pages · 5 pts | Source |
| A02 | Product content visible in a real browser Agents use a browser. The product name, price and buy button must actually render, not stay empty or spin. | Accessible site and browser-readable pages · 5 pts | Source |
| A03 | No challenge shown to our test browser CAPTCHAs, "checking your browser" screens and 403 walls stop automated browsers. ChatGPT's verified browser can be allowed through, but only if your security settings permit it. | Accessible site and browser-readable pages · 5 pts | Source |
| A04 | OpenAI search crawler rules in robots.txt OAI-SearchBot decides whether products can show up in ChatGPT search answers. It is separate from the browser that does the buying. | Not scored | Source |
| A05 | Verified ChatGPT browser access Wallet purchases happen in ChatGPT's cloud browser. It signs every request (Web Bot Auth, Signature-Agent "https://chatgpt.com"). If your CDN or firewall blocks it, the agent never reaches checkout. | Needs merchant verification | Source |
| A06 | CDN, firewall and bot-management providers detected Knowing which provider sits in front of the store tells you which setting to check. Detection is a clue, not proof of configuration. | Not scored | Source |
| B01 | Product pages can be found An agent asked to buy something has to find the right product page from your site, sitemap or catalog. | Product discovery and information accuracy · 5 pts | Source |
| B02 | Specific product names and descriptions Agents match what the shopper asked for against your product name and description. Empty or generic text causes wrong picks. | Product discovery and information accuracy · 5 pts | Source |
| B03 | Price and currency are clear and consistent The agent has to tell the shopper the price before they approve payment. A missing currency or a price that disagrees with your structured data is a trust failure. | Product discovery and information accuracy · 8 pts | Source |
| B04 | Stock availability is stated An agent should not try to buy something that is sold out. Availability needs to be stated, both on the page and in structured data. | Product discovery and information accuracy · 5 pts | Source |
| B05 | Product images load and have text alternatives Agents confirm "is this the right item" partly from images and their alt text. | Product discovery and information accuracy · 4 pts | Source |
| B06 | Variant options are labelled Sizes, colours and other options have to be readable and labelled, or the agent cannot pick the one the shopper asked for. | Product discovery and information accuracy · 3 pts | Source |
| B07 | Product structured data (JSON-LD) Structured data hands agents the facts directly: name, price, currency, availability, image. It is helpful, not a Wallet requirement. | Product discovery and information accuracy · 5 pts | Source |
| B08 | Product feed field readiness (local diagnostic) OpenAI accepts product feeds through Business Hub. We checked whether your public pages already contain the required fields. This is not a submitted feed. | Not scored | Source |
| B09R | Returns / refund policy published Shoppers approve agent purchases with less hesitation when the return terms are one click away. Agents look for them. | Fulfillment, policies and customer trust · 8 pts | Source |
| B09S | Shipping and delivery terms published The agent has to tell the shopper what shipping costs and how long it takes before they pay. | Fulfillment, policies and customer trust · 8 pts | Source |
| B09C | Contact, terms and privacy pages linked Terms and privacy URLs are fields OpenAI asks for when products are eligible for checkout; contact details are a basic trust signal. | Fulfillment, policies and customer trust · 4 pts | Source |
| C01 | Add-to-cart button present and usable This is the button the agent presses. It must be a real, visible, enabled control with a clear name. | Publicly observable cart and purchase pathway · 10 pts | Source |
| C02 | Cart entry point is discoverable After adding an item, the agent needs an obvious way to the cart and checkout. | Publicly observable cart and purchase pathway · 8 pts | Source |
| C03 | Purchase controls are machine-readable Agents act through the page's accessibility tree. Unlabelled divs pretending to be buttons, or unlabelled quantity boxes, cause misclicks. | Publicly observable cart and purchase pathway · 7 pts | Source |
| C04 | Cart item actually created Only an authorised test proves the add-to-cart action works end to end. | Needs merchant verification | Source |
| C05 | Cart persists across pages Carts that empty on navigation break the agent mid-purchase. | Needs merchant verification | Source |
| C06 | Guest checkout or supported sign-in Required accounts are the most common checkout blocker. OpenAI offers Sign in with ChatGPT for stores that need an account. | Needs merchant verification | Source |
| C07 | Shipping, tax and total shown before payment The shopper approves a total in ChatGPT; it has to match what your checkout charges. | Needs merchant verification | Source |
| C08 | Order confirmation returned A purchase only counts when the store creates the order and returns a confirmation. | Needs merchant verification | Source |
| D01 | Payment providers detected on public pages Wallet sends your existing checkout a network card token. Knowing your processor tells you who to ask about agent-initiated payments. | Not scored | Source |
| D02 | Wallet payment accepted Only a real, authorised Wallet purchase proves the payment step works. | Needs merchant verification | Source |
| D03 | Fraud and risk rules reviewed for agent traffic Risk systems often flag cloud browsers as bots by IP or browser signals, declining or cancelling good orders. | Needs merchant verification | Source |
| D04 | Forwarded user signals ingested OpenAI is rolling out a Forwarded header (asn, for, ip-country, ip-prefix, locale, tz, user-agent) so risk systems see the shopper, not the cloud browser. | Needs merchant verification | Source |
| D05 | Subscriptions and saved credentials Wallet credentials are single-use. Recurring billing needs explicit consent and your processor's merchant-initiated transaction flow. | Needs merchant verification | Source |
| D06 | Wallet order confirmed and not cancelled Good orders that your fraud rules cancel later still count as failures. | Needs merchant verification | Source |
| E01 | Business claimed in Business Hub Claiming the business lets you manage how it appears in ChatGPT. | Needs merchant verification | Source |
| E02 | Product feed submitted to OpenAI A submitted feed is how products are listed reliably. Schema on your pages is not the same as a submitted feed. | Needs merchant verification | Source |
| E03 | Plugin, MCP or WebMCP integration Optional. Structured tools (product lookup, cart, checkout handoff) make agent shopping more reliable. Never a Wallet prerequisite. | Not scored | Source |
| E04 | Agent orders measured Without tracking you cannot tell whether agent orders are happening or failing. | Needs merchant verification | Source |
| E05 | Platform recognised and page health Platform determines which fixes apply. Broken scripts and redirect chains slow or break agents. | Not scored | Source |