Agentic commerce
Buying where an AI agent does the research, and sometimes the checkout, on a person's behalf, so the product data a machine can read matters before the shopper ever sees a page.
For twenty years the shopper was the reader. They compared, they judged the photos, they decided the price was fair. Agentic commerce moves part of that job to software that reads your product data instead of looking at your page.
What agentic commerce means
Agentic commerce is buying in which an AI agent acts on a person's behalf: gathering options, comparing them, assembling a cart, and in some arrangements completing the purchase.
It runs on a spectrum. At one end, an agent researches and hands back a shortlist, and the person clicks through and buys as usual. At the other, the person states a constraint like a budget and a deadline, and the agent transacts once the conditions are met.
The common thread is that a machine consumes your product information before any human sees your page. What it can read is what it can act on.
Where the standards actually stand
Two named efforts are worth knowing, both hedged, because this area moves fast and program details change.
OpenAI announced Instant Checkout in ChatGPT on September 29, 2025, and open-sourced the underlying Agentic Commerce Protocol, co-developed with Stripe. At launch it covered single-item purchases from US Etsy sellers, with Shopify merchants described as coming soon, and merchants remaining the merchant of record for payment, fulfillment, and support. Source: OpenAI, "Buy it in ChatGPT," openai.com, September 29, 2025.
Google announced the Agent Payments Protocol, or AP2, on September 16, 2025, developed with more than 60 payments and technology organizations. Its core idea is a signed record of what the user actually authorized: an intent mandate capturing the request, and a cart mandate fixing the exact items and price the user approved. Source: Google Cloud blog, September 16, 2025.
Two qualifications matter more than the announcements. The Agentic Commerce Protocol site states plainly that implementing the protocol does not automatically list your products through AI agents, and that each AI platform runs its own participation process. OpenAI's commerce documentation describes product feed onboarding in ChatGPT as available to approved partners who apply. Both checked 2026-07-30.
So this is not a switch you flip. Eligibility is gated, the program details have changed since launch, and anything written about them today deserves a check against the platform's current documentation before you act on it.
What it changes about your product content
Strip away the protocols and the requirement underneath is old and familiar: a machine has to be able to read your product facts without interpreting a layout.
- Facts as data, not decoration. A price in a styled box and a stock status implied by a greyed-out button are visual conventions. Product schema states the same facts as labeled data a parser reads directly.
- Variants modeled individually. Where price, availability, images, or description differ by size or color, they need to be stated per variant with stable identifiers rather than collapsed into one parent record. OpenAI's own feed best-practice guidance calls this out explicitly, keeping title, URL, description, media, availability, and price variant-specific when they differ (developers.openai.com, checked 2026-07-30).
- Plain, factual descriptions. Marketing prose that gestures at a benefit gives a parser nothing to compare. Concrete attributes do.
- URLs and links that resolve. Product URLs and seller links have to be valid, encoded, and durable, because an agent follows them rather than eyeballing them.
None of this is exotic. It is structured data applied with more discipline than a store usually needs, because the reader is less forgiving.
Honest availability is the part that bites
Here is the failure mode specific to agents. A stale label is worse than a missing one, because it states a wrong fact with confidence.
If your schema says in stock and the item is not, or carries a price you no longer charge, a machine can repeat that to a shopper, or in a checkout flow act on it, while your visible page is perfectly correct. A human would have noticed the mismatch. A parser reads the label.
That makes availability and price accuracy an operational requirement, not a markup task. Feeds and schema have to track the catalog they describe, and they drift the moment nobody is watching.
What this does not mean
Being readable does not make an agent choose you. Selection is the agent's call, influenced by things outside your markup, and no tool can promise placement, recommendation, or a sale. Anyone selling that certainty is selling a number they do not control.
Readable product data also does not fix a weak offer, a bad price, or slow shipping. It removes the failure where good products lose because their facts were unreadable.
And to be specific about tooling: Citedon's scan works on any store, Shopify and headless setups included, and shows what engines cannot read on your product and category pages. The automated fix layer is WordPress-only, so elsewhere you get the diagnosis and apply the structure through your own theme or platform.
How to check yours
Take your best-selling product page and read only what a machine would get: the labeled facts, not the photography. Is there a price it can parse, a stock status it can trust, and a distinct record for the variant a shopper would actually order?
Run a free scan on that URL to see how engines read it today. For the wider picture on stores, read AI search for ecommerce stores, and for the markup itself, add product schema in WordPress.