Getting mentioned by AI isn't the same as getting bought from

Ask most "AI commerce" tools what they measure and the answer is whether ChatGPT or Perplexity mentions your brand. That's the wrong number to chase. A mention is discovery; the sale happens at transaction, and an agent that recommends you but can't complete a checkout on your store hands the order to whoever it can buy from instead.

Being found and being bought from drifted apart over the first half of 2026, and the gap is where stores quietly lose revenue.

The year the checkout got hard

OpenAI shipped Instant Checkout, then walked it back. By March 20 it had onboarded only about 30 Shopify merchants, ran into a sales-tax compliance wall, and pivoted to handing shoppers off to the retailer's own site to close the sale. Stripe co-authored the protocol and Shopify was the launch surface, and it was still harder than the announcement implied.

Walmart went the other direction. It dropped OpenAI's checkout entirely and plugged its own agent, Sparky, into ChatGPT as a plugin so it keeps control of the transaction. In May it reported Sparky users carrying a 35% higher average order value than non-Sparky shoppers.

BigCommerce crossed from discovery into transaction in late April with PayPal Store Sync: live agent checkout through Perplexity, Meta AI, and Copilot, with PayPal handling payment, fraud, and buyer protection. It ships as an app install rather than a default-on toggle, which tells you something about how many merchants it actually reaches.

Read those three together and the shape is clear. Discovery is broad and mostly automatic now. Transaction is where the work, the friction, and the money sit, and it's nowhere near solved.

Why "did AI mention me" points at the wrong thing

A citation tool tells you ChatGPT said your name. That's worth knowing for awareness and useless for the question your finance lead will ask, which is how many orders the agents actually placed.

An agent's decision to transact runs on different inputs than its decision to mention. Does the product carry a GTIN it can match against other sellers? Is price and availability current? Does the offer block parse? Can it read your return policy without hitting a 403? Is there a checkout path the protocol recognizes at all? None of that surfaces in a "share of AI voice" dashboard, and all of it decides whether the cart gets placed.

It's also why "isn't this just SEO?" misses. Ranking gets you seen. Whether an agent can buy turns on structured clarity, and that work overlaps with schema SEO without being the same job. More on that distinction in the agentic commerce guide.

The hedge that works no matter which protocol wins

You don't have to bet on a protocol. ACP runs ChatGPT shopping, UCP runs Google AI Mode and is what Copilot aligns to, Perplexity crawls your pages and reads their markup directly, and underneath the branding they all consume the same product data. Get that data right once and you're legible to every surface at the same time.

In rough order of return: complete schema.org/Product JSON-LD on every product page, GTINs on everything that has one, accurate real-time price and availability, crawlable shipping and return policies, AI crawlers allowed in robots.txt, and server-rendered pages because the bots don't execute JavaScript. An /llms.txt manifest on top is a cheap, honest signal.

None of that is a platform migration. It's data hygiene. It costs little and it pays off whether the surface that ends up driving your orders is UCP, ACP, or something that hasn't shipped yet.

So check the number that matters

If you've been watching whether AI mentions your brand, you've been watching the easy metric. The one tied to revenue is whether an agent can finish a purchase, and you can see where you stand: we score a store across all five surfaces against the live ACP and UCP specs and return a ranked, deep-linked fix list. Start with the full check list, or run the free audit on your own URL.

One honest caveat. We don't run a live test-buy on your store; that's still R&D and we won't pretend otherwise. What we check is readiness, the data and discovery files and policies and checkout signals the specs require. That's the line between guessing and knowing whether an agent can buy from you.