Merchant guide
The straight dope on agentic commerce: a catalog checklist for brands and merchants
A practical checklist for brands and retailers getting ready for agentic commerce: what is already handled at checkout, and the three things your product catalog still needs.
On November 13, OpenAI launched ChatGPT Wallet. A shopper approves an order, and an agent pays for it on the store’s normal checkout page, with a one-time card number from Visa, Mastercard or American Express. The store changes nothing.
That settles a question merchants have been asking all year: “what do I need to build so agents can buy from me?” For most stores the answer is now “very little.”
The harder question is the one before checkout. When a shopper asks an assistant for “a linen summer dress under $200,” the agent reads a handful of products and recommends two or three. Getting onto that list is the work, and it comes down to your product catalog.
This guide covers both halves: the short one (payment) and the long one (catalog), with a checklist at the end.
Part 1: Taking payment from an agent
You are probably covered already. There are three routes, and most stores are on one of them.
- You’re on Shopify. Shopify’s agentic storefronts list your products inside AI assistants and run the order through your Shopify checkout. Check that the channel is turned on in your admin.
- You take payments through Stripe. Stripe supports the Agentic Commerce Protocol, where the assistant passes a single-use payment token to the Stripe integration you already have.
- Neither. ChatGPT Wallet works on your existing checkout page. The agent fills in the form the way a shopper would, using a card token that is good for that one order.
What is left for you to do:
- Talk to your fraud and bot-protection vendors. An agent filling in a checkout form can look like an attack. OpenAI says so itself: risk systems may decline legitimate agent orders. Ask your vendor how it recognizes agent checkouts and what you need to switch on.
- Decide on purpose whether to let agents in. Some retailers block them. That is a real choice, but it should be a choice. Check your
robots.txtand your bot-protection rules for a default that blocks ChatGPT, Claude and Perplexity without anyone having decided it.
That is the payment half. It is mostly a settings review.
Part 2: The catalog is where the work is
We ran AI visibility audits on ten established direct-to-consumer brands. Each audit asks ChatGPT, Claude, Perplexity and Google the questions a shopper would ask without naming a brand, and counts how often the store shows up.
Roughly half of the ten did not appear in a single one of 48 answers.
These are not broken websites. On crawlability (sitemaps, metadata, basic structured data) they score well. On agent readiness, which measures whether an agent can understand and compare the products, all ten scored weak. A site can be easy to crawl and still hard for an agent to read.
An agent meets your catalog in three ways. Each one needs something different from you.
1. Make every product understandable
An agent that opens one of your product pages needs eight questions answered:
- What is it? A product type, not only a name. “The Sunday” is a name. “Linen midi wrap dress” is an answer.
- What does it look like? Color, pattern, silhouette, and images it can reference.
- What is it made of? “55% linen, 45% cotton,” not “a breezy blend.”
- Will it fit? Sizes offered, the cut, and measurements.
- What does it cost, and can I buy this size in this color today?
- Who is it for? Gender, age group, occasion.
- Is it any good? A rating and a review count.
- What makes it different from every other item in its category?
Most product pages answer these for a person. The answers sit in photos, in a details accordion, in a size-chart popup, in a reviews widget. An agent reading the page’s text and markup gets some of them and guesses the rest.
Two tests are worth running on your own products:
- Could the agent pick it out of a lineup? “Brown multi” does not pass. “Brown houndstooth with an antique key print” does.
- Would it be found? Shoppers do not search for “The Sunday.” They ask for linen, midi, wrap, wedding guest, under $200. If those words are not attached to the product as attributes, the product is not a candidate.
Then there is cost. A typical product page is tens to hundreds of thousands of tokens of markup and scripts. Some pages are an empty shell until JavaScript runs. An agent on a budget reads less of you, or reads a competitor.
Check that each value is in the right field
This one is easy to miss, because the page looks fine to a shopper. Merchandising has been manual work, so values end up in fields that were never meant for them. We routinely see a fabric in the brand field, or a season label attached to the brand name (“Brand - SS25”). On Shopify the usual source is the vendor field. It becomes the brand in feeds and structured data, and it often holds a supplier, a collection or a season instead.
A shopper never sees these fields. An agent filters on them. Ask for dresses from your brand, and the ones filed under “Linen” or “Brand - SS25” are not in the results. Every channel downstream inherits the mistake: your structured data, your feed, and any search built on the same fields.
What to do: put material, pattern, fit, audience, variants and ratings into your product structured data, not only into the page design. Publish a plain Markdown version of each product page and link to it from the page. Then export the distinct values of each field (brand, material, color, category) and read the list. Values in the wrong field stand out quickly.
How StoreRouter helps: Agent Mirror does all three. We already read your store, classify each product, and add the attributes agents filter on. Along the way we check that each value fits the meaning of its field: a brand field holds brands, a material field holds materials. Misplaced values are corrected, and real partner brands and collaborations are kept as they are. Agent Mirror then serves every product as a clean Markdown page of about a thousand tokens, on a subdomain of your choice. You add one link tag to your product template.
2. Make the whole catalog available
Assistants do not want to discover your products one page at a time. They want the full list, kept current.
ChatGPT now accepts product feeds directly from merchants. You claim your business, verify your identity, and deliver a feed by upload, hosted URL or SFTP. It is US-first for now. Expect the other assistants to ask for the same thing.
The feed has a short list of required fields (title, description, price, availability, image, brand) and a longer list of optional ones: material, color, size, gender, age group, category, dimensions, variants, ratings. Being in the feed does not guarantee being shown. The optional fields are what let the assistant match your product to a specific request, so they are optional in name only.
A feed also copies whatever is in your fields. If the brand field holds a fabric or a season label, that is what the assistant receives, and brand is one of the required fields.
What to do: fix misplaced values first, then submit a feed. Fill in every attribute field you can. Update it whenever prices, sales or stock change.
How StoreRouter helps: the same enriched catalog behind Agent Mirror can be published as a feed, with the attribute fields filled in and kept fresh. We are building this now: one hosted feed URL that you hand to each assistant.
3. Make your catalog searchable
The third way in is live. An agent asks your store a question (“gold-frame sunglasses with a pink tint”) and expects products back.
The emerging standard is the Universal Commerce Protocol. A store publishes a small file at /.well-known/ucp that tells agents what it supports, including catalog search. Shopify generates one for its stores. You can also point agents to a search endpoint from your llms.txt or agents.md.
Having the endpoint is the easy part. A search endpoint can only filter on the attributes it has. We tested this against Shopify’s catalog search with seven attribute-heavy fashion queries, judged blind. Where Shopify has filters (category, gender, color), the two engines tied at 99%. On the attributes it cannot filter on (pattern, fabric, fit, finish), results matched 98% of the time on StoreRouter and 73% on Shopify. For “glossy women’s shoes with buckles,” one of Shopify’s top ten matched. Shopify did win one of the seven queries.
What to do: check whether your store answers at /.well-known/ucp. Then test it the way a shopper would talk: ask for a fabric, a fit, a finish, and see what comes back.
How StoreRouter helps: StoreRouter’s search runs on the enriched attributes, so it can filter on the things shoppers actually ask for. Agent Mirror will expose that search for your store alone, at an endpoint agents can call.
The checklist
| What to check | If it’s missing | |
|---|---|---|
| Payment | You are on Shopify or Stripe, or your checkout works with ChatGPT Wallet | Usually nothing to build |
| Your fraud and bot tools let approved agent orders through | Ask your vendor | |
robots.txt does not block answer engines by accident | Edit robots.txt | |
| Product pages | Each product answers the eight questions in text and structured data | Agent Mirror |
| Each field holds the right kind of value: no fabrics or seasons in the brand field | Agent Mirror | |
| Each product has a Markdown version linked from the page | Agent Mirror | |
| Catalog | A feed is submitted to ChatGPT with attribute fields filled in | Agent Mirror feed (coming) |
| Search | Agents can search your catalog by fabric, fit, pattern and finish | Agent Mirror search (coming) |
Start by seeing what agents see
Before changing anything, find out where you stand. Our free audit asks the major assistants real shopping questions in your category, shows who they recommend instead of you, and scores your product pages on the same things this guide covers.
Then turn on Agent Mirror. It is free to start, and we have probably already indexed your store.
— Mahmoud