Agentic commerce is when an AI assistant finds a product, compares options and completes the purchase on a shopper’s behalf. Three competing protocols now govern how that works, and the one most articles tell you to chase is the one that failed. ChatGPT’s Instant Checkout ran for roughly five months and was shut down in March 2026 with near-zero sales. Google’s Universal Commerce Protocol is the one actually transacting today. Here is what that means for your store.
The three protocols, and what each one does
They are less rival than complementary, though they overlap at checkout.
Protocol | Job | Backed by | Status |
|---|---|---|---|
UCP — Universal Commerce Protocol | Discovery, cart, checkout, post-purchase | Google, Shopify, Etsy, Wayfair, Target, Walmart, plus 20+ endorsers including Visa, Mastercard, Amex, Adyen, Stripe | Live. Announced Jan 2026 at NRF |
ACP — Agentic Commerce Protocol | Executes checkout inside an AI surface | OpenAI and Stripe | Consumer product withdrawn; survives as infrastructure |
AP2 — Agent Payments Protocol | Proves a human authorized the purchase | Google, donated to the FIDO Alliance April 2026 | v0.2, authorization layer |
UCP works by having merchants publish a manifest at a well-known URL and support at least one transport. ACP’s core mechanism is a shared payment token scoped to one merchant and one basket total, so the agent never sees card details. AP2 uses three signed mandates covering intent, cart and payment.
What actually happened to ChatGPT checkout
This is the part worth knowing before you spend anything.
OpenAI launched Instant Checkout with Etsy and said a million-plus Shopify merchants were coming. Reporting by Digital Commerce 360 indicates the consumer product was withdrawn in March 2026 after about five months, with near-zero sales, and that only around a dozen of the eligible Shopify merchants ever integrated it.
ACP itself did not die. It survives as plumbing: Target, DoorDash and Instacart run ChatGPT Apps that transact through it. But the vision of shoppers buying directly inside a chat window, which drove most of the coverage in late 2025, did not convert.
Two lessons sit in that. Protocol announcements are not demand. And ChatGPT having hundreds of millions of weekly users turned out not to mean those users wanted to buy inside it.
What is actually live
Google’s UCP checkout is running for eligible US retailers across Search and the Gemini app, accepting Google Pay and PayPal. Business Agent on Search is live with Lowe’s, Michael’s, Poshmark and Reebok. A Direct Offers pilot includes Petco, e.l.f. and Samsonite.
Shopify merchants got a meaningful unlock in June 2026: any developer can register an agent profile in the Developer Dashboard and call the public endpoint with no approval gate. Google added cart transfer to a merchant’s own site in September 2026, rolling out in the US first with Australia and Canada expected in early 2027.
OpenAI’s feed onboarding, meanwhile, remains limited to approved partners who apply through a form.
Is any of this worth your time yet?
The numbers point in two directions and both are worth holding.
Shopify’s own Q1 2026 data reports AI-referred orders growing nearly 13 times year on year, with AI-referred visitors converting at close to 50% higher rates than organic search. Gartner forecasts that by 2030, 20% of transactions will run through AI platforms. Shopify also claims searches using its Catalog API convert at twice the rate of scraped data, though that is a vendor claim with no independent audit.
Against that: the absolute volumes are still small, one flagship consumer product has already failed, and the protocol landscape has changed three times in eighteen months.
The sensible read is that the growth rate is real and the base is low. That argues for cheap, protocol-agnostic preparation rather than expensive integration work tied to any one standard.
The five things worth doing now
All of these help every protocol, and all of them also improve ordinary shopping feeds and conventional SEO. That is the test a task should pass before it gets your budget.
- Find out where your product data actually lives. If price, availability or specification only appears after JavaScript runs, or exists only inside a theme template or display rule, software cannot query it. Agents then fall back on scraping your pages, badly. This audit is the first job and it is usually the one that finds the real problem.
- Group your variants under one parent. A product with five colours should read as one product with options, not five separate products. Separate pages per variant make an agent treat them as unrelated items and dilute everything.
- Use the most specific taxonomy you can. “Men’s insulated winter boots”, not “footwear”. Specific categorisation is how an agent decides whether your product answers the question it was asked.
- Write product fields literally. “Ocean Breeze” is a lovely product name and it tells a machine nothing. “Texturizing sea salt spray” does. Keep the marketing voice in your description copy and keep the product fields plain.
- Solve real-time price and inventory. Agents want accuracy at the moment of the query, not at your last crawl. Serving this through a real-time endpoint is the most resource-intensive item on this list, and building a separate one per platform is genuinely hard. Start by making sure your existing feed refreshes often rather than building endpoints for protocols you have not committed to.
What each platform currently asks for
Requirements differ, and this is where integration costs live.
Platform | What it requires |
|---|---|
Google / Gemini | Enriched feed to Google Merchant Center, augmented with GenAI attributes; host a UCP manifest file; build a real-time inventory check endpoint |
OpenAI / ChatGPT | Apply as a merchant partner; build a product feed to OpenAI’s specification (CSV or JSON, with daily refresh) |
Perplexity | Integrate via Feedonomics, the Sonar API, or their enterprise resource hub |
One detail worth noting on the OpenAI route: the merchant remains the merchant of record. You validate the order, calculate and charge sales tax, screen for risk and fraud, charge the payment method, and accept or decline. OpenAI renders the interface. The liability stays with you.
Where to start if you only do one thing
Fix your product data. It is the only work on this list that cannot be wasted.
Every protocol reads the same underlying attributes. Clean variant structures, specific taxonomy, literal field language and accurate availability improve your Google Shopping performance today, your conventional search visibility today, and your position with whichever protocol wins. Integration work does not carry that guarantee, as the dozen Shopify merchants who built for Instant Checkout discovered.
The honest summary
If you sell online in the US, publish a clean enriched feed to Google Merchant Center and get your product data into a machine-readable state. That is proportionate to the current opportunity.
If somebody is selling you a bespoke agentic commerce integration in 2026, ask which protocol, how many transactions that protocol processed last quarter, and what happens to the work if it goes the way of Instant Checkout.
We will tell you where your store actually stands
We will audit your product data, check what an agent can and cannot read from your store, and give you a prioritised list. No protocol commitments required.
No charge, and the findings are yours either way.
Frequently Asked Questions
What is agentic commerce?
Agentic commerce is when an AI assistant completes a purchase on a shopper’s behalf — finding the product, comparing options, and checking out without the person visiting the store. Three protocols govern it: UCP for discovery and cart, ACP for checkout execution inside an AI surface, and AP2 for proving the human authorized the transaction.
Is ChatGPT Instant Checkout still available?
The consumer product was withdrawn in March 2026 after roughly five months, reportedly with near-zero sales and only around a dozen eligible Shopify merchants integrated. The underlying protocol survives as infrastructure — Target, DoorDash and Instacart run ChatGPT Apps that transact through it — but buying directly inside a chat window did not take off.
Which agentic commerce protocol should merchants support?
UCP has the clearest current adoption. Google’s UCP checkout is live for eligible US retailers on Search and the Gemini app, with Business Agent running with Lowe’s, Michael’s, Poshmark and Reebok. For most merchants the practical answer is a clean enriched feed to Google Merchant Center rather than a bespoke integration with any single standard.
How do I prepare my store for AI shopping agents?
Five things: audit where your product data actually lives and move anything trapped in JavaScript or theme templates; group variants under one parent product; use the most specific taxonomy available; write product fields in literal language rather than marketing names; and make price and inventory accurate in real time rather than at last crawl.
Is agentic commerce worth investing in yet?
The growth rate is real and the base is low. Shopify reports AI-referred orders up nearly 13x year on year with conversion around 50% higher than organic search, and Gartner forecasts 20% of transactions running through AI platforms by 2030. That justifies cheap protocol-agnostic preparation, not expensive integration tied to one standard.
Who is responsible if an AI agent order goes wrong?
You are. On the OpenAI route the merchant remains the merchant of record: you validate the order, calculate and charge sales tax, screen payment and risk, and accept or decline. The AI platform renders the checkout interface. The liability does not transfer.