Agentic Commerce: Prepping 5 Shopify Stores for ChatGPT

Agentic commerce in ChatGPT from a five-store Shopify operator — the eligibility pass, product data cleanup, my checkout call, and what really showed up.

Linh Nguyen · Updated 2026-07-04

In the last week of March 2026 I blocked out two mornings to answer one question across the five Shopify stores we operate: if a shopper asks ChatGPT for the kind of products we sell, do we show up? Shopify had switched on Agentic Storefronts for eligible US merchants on March 24, so on paper the answer was already yes. In practice, "eligible" was doing a lot of work in that sentence, and I wanted to verify it with my own eyes in each admin.

This post is the prep log from that exercise — what agentic commerce in ChatGPT actually demanded from us as a Shopify operation: the eligibility pass, the product data cleanup that ate two weekends, the decision about where checkout happens, and what has shown up in our analytics since (honestly: not much yet). If you want the concept — how a storefront changes when the buyer is an AI agent — that lives in our agentic storefronts explainer, and the full technology map is in AI agents in Shopify 2026. This piece stays on the to-do list.

The dead end I almost prepared for: buying inside ChatGPT

I'll start with the mistake I nearly made, because dodging it reshaped my whole checklist. When OpenAI and Stripe launched "Buy it in ChatGPT" in September 2025 — entire purchases completed inside the conversation, running on the Agentic Commerce Protocol they maintain as an open standard — my first instinct was to prep our stores for in-chat transactions.

I never got past reading the docs, and I'm glad. Fewer than 30 of Shopify's millions of merchants ever went live with it, and Walmart reported conversion roughly three times lower than traditional checkout — Modern Retail's post-mortem is worth your ten minutes. Shoppers browsed and researched in the chat, then wanted to pay on a brand site they trusted. By March 2026 the in-chat payment experiment was effectively over: since March 4, 2026, Instant Checkout routes the shopper to the merchant's own storefront — an in-app browser inside ChatGPT or a new tab — instead of settling payment in the conversation. A handful of giants (Target, Sephora, Nordstrom, Lowe's, Best Buy, The Home Depot, Wayfair) run their own custom apps in ChatGPT, and PayPal is bringing its merchant network onto the platform through 2026, but for stores our size the model is now simple: discovery happens in the chat, the sale happens on your site.

That reframing set every priority that followed. I stopped thinking "new checkout integration" and started thinking "new front door" — which turns out to be mostly a data-quality job.

The eligibility pass: five admins, one embarrassment

Shopify's side of this shipped in December 2025 as Agentic Storefronts in the Winter '26 Edition: your products syndicated to ChatGPT, Microsoft Copilot, Perplexity, Google AI Mode and Gemini through Shopify Catalog — no app to install, no 4% transaction cut, no per-platform integration. The March 24 activation then flipped it on by default for eligible US merchants.

So my first task wasn't building anything; it was verifying. Store by store: is the store actually eligible, is Catalog picking up the products, and does the product data meet the basic standards agents get fed? Most of our five passed without drama. The embarrassment was the store I'd have bet money on — our oldest catalog, where early products still carried supplier-pasted descriptions and a batch of variants had no images at all. None of that blocks a human shopper landing from an ad. All of it degrades what an AI agent has to work with, because the agent only knows what the data tells it.

Two weekends of cleanup nobody will ever tweet about

Over two weekends in April 2026 we did the unglamorous part. The working rules I ended up with:

• Titles get read aloud now. A title stuffed with SKU codes and "[HOT]" tags was tolerable when it only had to rank; an agent quoting it to a customer in a sentence is a different test. We rewrote the worst offenders into plain product names.
• Deduplicate descriptions. Near-identical supplier text across similar products gives an agent no way to tell them apart — or a reason to recommend either.
• Inventory has to be genuinely live. Agents query real-time availability; stale stock counts turn into overselling faster than browser traffic ever did.
• Every variant needs an image and a price that isn't an accident. Obvious, and still where we found the most rot.

This is Catalog hygiene more than it is a ChatGPT feature — the long version is in our Catalog and product data quality deep dive, and GEO and llms.txt covers the discovery-optimization layer that sits on top.

One genuine dead end from those weekends: I started hand-polishing every product on the biggest store, burned most of a Saturday on maybe a tenth of the catalog, and gave up. Round two, we sorted by last-90-day revenue and cleaned top sellers first. If an agent ever recommends us, it will almost certainly be through those products — that's my estimate, not a measurement, but I'd stake the ordering of the work on it.

Where I let checkout happen (and where I didn't)

Shopify gives you the controls here: toggle which AI platforms you appear on, and choose whether a sale completes in the AI surface where supported or routes to your storefront. Orders land in your Shopify admin with AI channel attribution either way.

I route everything to our own storefronts, and it wasn't a hard call. Our payment setup and fraud checks live there with years of history behind them. And I already run one AI agent against these stores daily — since February 2026 it drafts support replies and runs checks, and it has never been allowed to touch money (the full delegation rules are in what I let an AI agent manage). I apply the same conservatism to other companies' agents: checkout rails I already audit beat a new payment surface I don't. If agent-driven volume ever justifies revisiting that, I'll revisit. It does not yet.

What the analytics actually show — the honest part

The platform-level numbers are genuinely large. Shopify reports AI-generated traffic to stores up 8x year over year and orders from AI-powered searches up 15x since January 2025 — their figures, aggregated across millions of stores, so any single store's mileage will differ. The March activation made roughly 5.6 million Shopify stores discoverable inside ChatGPT and its reported ~1 billion monthly users — again Shopify's and OpenAI's counts, not something I can verify from an admin screen.

Ours: since late March I've been watching for agent-referred sessions across the five stores, and there is a trickle — enough to prove the pipe exists. But I have yet to see agent-driven orders in any volume worth reporting. Small sample, one operator, a rollout that's a few months old — and exactly why I'd distrust anyone selling you a confident conversion benchmark for this channel in mid-2026. What our own data does tell me: being ready cost two weekends and some checklist discipline. Being unready gets expensive precisely when the channel arrives, because agents will recommend the clean catalogs that are already there.

The standards layer I'm watching but not acting on

Behind all of this, Shopify and Google are building the Universal Commerce Protocol (UCP) — an open standard for how agents handle cart, checkout, payment and post-purchase across any platform and processor, with founding members Google, Shopify, Etsy, Target and Wayfair, Amazon, Meta, Microsoft, Salesforce and Stripe on the UCP Tech Council, and endorsements from American Express, Mastercard and Visa. It's clearly designed to fix what the first checkout experiment got wrong: merchants keep direct control.

My action item as an operator, for now: nothing. Standards will consolidate on their own schedule. Your catalog you have to clean yourself.

Doing this five times over is the actual job

Every step above ran five times: five eligibility checks, five catalogs to keep clean, five analytics views to watch for agent-referred sessions, five places for attribution to live. That overhead — not any single toggle — is the real cost of agentic commerce for a multi-store operation, and it is the argument for consolidating orders, revenue and channel attribution into one view before you add the channel: a new channel should mean one new column, not five new logins.

My closing position, for whatever a sample of five stores is worth: agentic commerce ChatGPT Shopify integration is a discovery channel, not a checkout revolution — the checkout revolution was tried and shoppers declined it. Discovery channels reward whoever is prepared when the traffic shows up. Preparation, it turns out, costs two weekends.