Your Store Just Got a Second Storefront
- Shopping agents can search product catalogs, build carts and hand shoppers to a retailer's checkout.
- Product names, prices, stock, sizes and shipping rules need to be available as readable text.
- Anthropic's reported sales results came from one partner and have not been independently checked.
- I measure AI visibility with fixed questions, repeated checks and records that can be compared over time.
A shopping agent can search a catalog, build a cart and send the shopper to a retailer's own checkout inside one conversation. That gives a store a second storefront: the catalog information an agent can read.
On September 2, 2026, Anthropic released an open-source blueprint for shopping and merchant agents on Claude. Shopify published open-source code running that blueprint against a real Shopify store within about 48 hours. Digital Commerce 360 reported on Anthropic's release.
The practical question for a merchant is whether an agent can find the information a shopper needs. That starts with the product data, and it continues with checks that show whether the information is being found.
The takeaway
Treat the catalog as information an agent must be able to read. Check that it can find the facts a shopper needs before relying on it to sell.
What does a shopping agent do?
A shopping agent searches a catalog, builds a cart and hands the shopper to the retailer's own checkout, all inside one conversation. Anthropic's blueprint also includes a merchant agent for sales analysis, inventory alerts, pricing advice and campaign drafts. A person signs off before any change goes live.
The merchant agent's approval step sets a clear boundary: it can prepare advice and drafts, while a person approves changes. Before building an agent workflow, map who reviews its work and who can approve a change.
According to Anthropic, one partner reported that cart size grew about 30 to 35 percent and shoppers were 60 percent more likely to finish a purchase. These numbers reflect that partner's reported results. No independent review has confirmed them, and they don't show what another store can expect.
What product data can an agent read?
An agent can sell only what it can read. Product names, prices, stock, sizes and shipping rules need to be available as text it can find. If a shopper's decision depends on one of those details, leaving it out of readable product information can keep the agent from using it.
That makes catalog content part of how the store works for an agent. A person can scroll through pages and images, but the agent needs product facts in text it can locate. When details live in PDFs or images, an agent may not be able to quote them.
A practical review starts with a product page and asks whether the written information gives an agent the facts needed to find and describe the product. Check that prices, available stock, sizes and shipping rules appear as text. A gap in any of those fields leaves the agent with less information to use when helping a shopper.
How can I tell whether AI tools find my business?
I measure AI visibility by asking the same library of real buying questions of ChatGPT, Claude, Perplexity, Gemini and Grok on a fixed schedule. I check official domains so a company with the same name does not count as a win. I also make the model explain its choice and keep every result in one format, so week 11 can be compared cleanly with week 1.
The fixed questions make it possible to compare answers over time rather than rely on a single prompt. Checking the official domain separates a real mention of a company from a name match. Asking the model why it chose an answer gives me a record of its explanation, alongside whether it named the business.
During a separate program, I published in the content gaps, but the win rate did not change. In one engagement, Soapbox's AI recommendation win rate rose from 6.2% to 36.2% in 11 weeks. Trinzik carried out that work using the platform I built. Treat this as directional evidence, not a controlled study. I retest rather than call it a win.
I also published a rubric for scoring how readable a site's information is to agents. Any firm can score itself with the rubric and get the same answer. It gives a merchant a way to assess the site's readability alongside repeated checks of what AI tools recommend.
Common questions
Does the reported sales increase predict what my store will see?
No. Anthropic's figures came from one partner, were reported by Anthropic and have not been independently checked. Treat them as a reported result from that partner, not as a forecast for your store.
Does the shopping agent complete checkout?
The agent builds the cart and hands the shopper to the retailer's own checkout inside the conversation. For your own review, check where that handoff happens and whether your checkout remains the place where the shopper completes the purchase.
Can I assess agent readability without hiring a consultant?
Yes. I published a rubric that any firm can use to score its site's readability by agents and get the same answer. Use that score to find information that may be hard for agents to read, then check whether the tools you track can find your business in repeated tests.
Sources
If you want help checking what AI tools can read and recommend about your store, contact me.