Being Cited by AI Isn't the Finish Line Anymore — Being Bought by It Is
For a few months last year, "sell through ChatGPT" meant one specific thing: Instant Checkout, launched with Stripe in September 2025, let someone buy an Etsy listing without leaving the chat window. By March 2026, OpenAI had quietly pulled it back. Only a handful of merchants had bothered to integrate. The headline write-up was "agentic checkout failed." The more useful read is narrower: that one feature failed. The thing underneath it didn't.
The checkout experiment that didn't work — and the layer that did
While Instant Checkout was stalling out, Shopify was doing something that mattered more: it switched on Agentic Storefronts by default for every eligible store, syndicating product listings into ChatGPT, Google AI Mode, Microsoft Copilot, and Perplexity at once. Nobody had to request it. A store simply became buyable through AI surfaces it was never built for, running on protocols — Google and Shopify's Universal Commerce Protocol, OpenAI's now-open-sourced Agentic Commerce Protocol, Visa's Trusted Agent framework — that most merchants have never read a line of. The checkout button was the part that got attention. The product-feed plumbing underneath it is the part that actually shipped.
Discovery, not checkout, is where the real contest is happening
After the Instant Checkout retreat, ChatGPT leaned harder into comparison and recommendation instead of completing the sale itself — which tracks with where the trust actually sits. Research from Bain in 2026 found people trust a retailer's own on-site AI agent roughly three times more than a third-party one doing the buying for them. That's not a technology gap, it's a confidence gap: someone will happily ask an AI what to buy, and still wants to land on a site they recognize to actually buy it. The agent's job is narrowing the field to two or three options. Getting into that shortlist is the new "page one."
What actually makes a product legible to an agent
The same discipline from getting cited in AI answers applies here, one layer deeper. An agent comparing products needs structured, unambiguous data — price, availability, specs, reviews — exposed the way these new protocols expect it, not buried in a product page built for a human scrolling with a mouse. It needs that data to say the same thing everywhere: a product described one way on a site and another way in a feed reads to a model as two different, less trustworthy products. And none of it matters if the underlying site is slow or badly structured enough that an agent can't crawl it reliably in the first place — the same foundational problem that quietly sinks a site that looks finished but fails Core Web Vitals.
The part no protocol can hand you
Here's what the whole agentic-commerce conversation keeps skipping: an agent can be fed perfectly structured data and still have nothing distinct to say about a product, because clean data isn't the same as a reason to choose one option over the other three in the shortlist. That reason was never going to come from a protocol. It comes from the same place it always has — a product or a brand that's actually different, stated clearly enough that a machine can repeat it back correctly.
This is a Grow post — the last of our four stages: Define, Express, Build, Grow. If your products are technically "in" every AI shopping surface but never the one an agent actually recommends, Grow is where that gets fixed.
