Every "optimize for AI shopping" article landing in your inbox this month is already behind the news. Most of them are still written as if ChatGPT Instant Checkout, the feature that let shoppers buy from an Etsy seller without leaving the chat window, is the thing to prepare for. OpenAI effectively retired it in March 2026, about six months after launch. Walmart's own EVP of product and design said purchases completed inside ChatGPT converted at one-third the rate of shoppers who clicked through to Walmart.com, and called the in-chat experience "unsatisfying." The feature that made headlines a year ago is functionally gone. The protocol underneath it is not, and that distinction is exactly what most AI shopping agents advice gets backwards.
AI shopping agents are not a future trend to prepare for. They are already reading your product feed, and in a growing number of cases, already completing checkouts on behalf of your customers. The open question is not whether to care. It is whether your tracking, attribution, and product data are actually built to survive contact with them.
- OpenAI killed in-chat checkout in March 2026 but kept the Agentic Commerce Protocol, co-developed with Stripe, as the open standard for how agents discover products and hand off to merchant checkout.
- Five separate AI shopping surfaces, ChatGPT, Perplexity's Instant Buy, Google's Universal Cart, Copilot, and Alexa, are converging on the same underlying requirement: a complete, accurate, real-time product feed, not a content strategy.
- The readiness gap nobody is measuring: what happens to your GA4 purchase event and channel attribution when the shopper never lands on your site the normal way.
What actually happened in the first year of agentic commerce
The timeline moved fast enough that most "2026 AI shopping checklist" content is already describing a version of the landscape that no longer exists.
OpenAI and Stripe released the Agentic Commerce Protocol as an open standard in late 2025, alongside Instant Checkout, the feature that let ChatGPT complete a purchase without sending the shopper anywhere. It shipped first with Etsy.
By March 2026, OpenAI had pulled the in-chat checkout experience and repositioned ChatGPT around discovery, with the actual purchase happening back on the merchant's own site or inside a merchant-built app using OpenAI's SDK. The protocol itself, now maintained as an open specification with Stripe, kept going. The checkout hand-off model still needed a standard way for an agent to pass cart contents, delegated payment tokens, and order status back and forth with a merchant, regardless of which chatbot initiated it.
Google followed a different path to a similar place. At I/O 2026, Google introduced Universal Cart, which collects items a shopper adds from Search, the Gemini app, YouTube, and Gmail into a single cart that can track price drops and compare checkout options across merchants in the background. Perplexity took a third route, partnering with PayPal to launch Instant Buy in November 2025, letting US shoppers complete a purchase inside Perplexity itself for participating merchants, PayPal remaining the merchant of record.
Three different companies, three different technical approaches, and one shared conclusion: discovery is moving into the AI interface, and checkout is either happening there too or handing off cleanly to the merchant. Either way, the product data an agent reads has to be right before any of it matters.
The readiness advice that skips the part that actually breaks
Most "is your store ready for AI shopping agents" content converges on the same list: Schema.org Product markup, complete GTINs, accurate real-time pricing and availability. That list is correct as far as it goes. Google's own structured data documentation confirms Product markup has to be present in the HTML returned by the server, not injected by JavaScript after the page loads, because it functions as a verification layer against your Merchant Center feed. When the two disagree, Google deprioritizes both, and the same feed increasingly powers more than just a shopping ad.
Feed accuracy is not a new problem AI agents invented. It is an old problem that now costs more. In a delivered paid media audit, we found roughly 11% of a client's catalog silently disapproved inside Google Merchant Center, the same product feed that powers advertising, Shopping, and now AI shopping surfaces alike. The cause was a currency app update that broke the price match between the feed and the live landing pages, plus missing GTINs on a subset of items. The disapproved SKUs included several best-sellers, and nobody on the team knew, because a disapproval does not throw an error on the storefront. The product simply stops appearing wherever the feed is read from, whether that is a Shopping ad or, now, an AI agent comparing options for a shopper.
Google's own data specification for Merchant Center treats GTIN as one of the strongest matching signals for connecting a listing to a shopper's query, and a wrong or missing GTIN is exactly the kind of gap that will not throw an error, only silence. BigCommerce's March 2026 pulse survey of ecommerce leaders found 40% of respondents were still in the process of standardizing product pages for agentic AI and 33% had not started at all, with none reporting their product data as fully optimized. The gap is not a knowledge problem. It is the same feed hygiene problem stores have carried for years, now with a second, faster-growing audience reading it.
The blind spot the feed checklists don't cover: your own attribution
Here is the part almost nobody selling an "agentic commerce readiness checklist" mentions: fixing your feed does not fix your tracking, and the two are not the same project.
When ChatGPT hands a shopper back to your own checkout, the redirect model that replaced Instant Checkout, your GA4 purchase event fires the same way it always has, assuming the tag is configured correctly in the first place. What changes is attribution. If the agent's browser or app strips the referrer, and a meaningful share of AI traffic already does this, that order lands in Direct instead of getting credit for the AI assistant that actually drove it. We covered exactly how that gap shows up and how to build a channel group that catches it in a separate piece on AI referral traffic in GA4, and the same mechanics apply to a completed purchase, not just a session.
The harder case is the app model, where a purchase completes inside the agent's own interface through a delegated payment flow defined by the Agentic Commerce Protocol. That transaction may never generate a page view or a purchase event on your site's own GA4 property at all, because the shopper's browser never rendered your checkout page. Revenue lands in your order management system and your payment processor's ledger. Whether it lands anywhere in your analytics depends entirely on whether someone built a reconciliation path between those systems and GA4, and for most stores today, nobody has. That is the same class of problem our ecommerce tracking audit exists to catch: not just whether a tag fires, but whether every path an order can take actually reaches your reporting. It is a fixed-scope engagement priced from $500, not an open-ended retainer, because the question it answers is specific.
What to actually do this quarter
Given how fast the last twelve months moved, chasing a checklist built around whichever single product is trending this month is a poor use of a quarter. Three moves hold up regardless of which agent wins the next round:
- Audit your Merchant Center feed and on-page Product schema together, not separately, since Google explicitly checks one against the other. Fix GTIN gaps and price mismatches before they silently drop best-sellers out of every surface reading that feed.
- Confirm your GA4 purchase event survives a redirect from an external AI interface, and build the channel-group logic to catch it correctly instead of letting it default to Direct.
- Ask whether any checkout flow in your stack could complete outside your own site entirely, and if the answer is yes or might soon be, get a reconciliation path between that system and your reporting before it happens, not after finance asks why revenue does not match sessions.
If the honest answer is that your team could also use a broader read on where AI actually saves hours internally, that is a related but separate question from agentic commerce readiness, and our piece on what an AI readiness assessment actually measures draws the line between the two so you are not buying the wrong audit for the problem you actually have.
Where this leaves you
Instant Checkout came and went in six months. Universal Cart and Perplexity's Instant Buy are still early enough that either could follow the same path, or become the default. Betting a quarter of engineering time on any single agent's current implementation is a reasonable way to end up rebuilding it by spring. Betting on feed accuracy, schema correctness, and tracking that survives a hand-off is not, because every version of agentic commerce that has shipped so far depends on all three regardless of which company's name is on the interface.
