We sent an AI shopping agent into 1,102 Shopify stores

AI agents are already shopping. Shopify switched on its agent commerce layer, the Universal Commerce Protocol, for stores this year, so a ChatGPT- or Claude-style agent can find a product, put it in a cart and hand the shopper a checkout link without ever opening a browser.

We wanted to know whether that works in practice, so we tested it: 272 of the best-known Shopify brands and 830 mid-size merchants pulled from Shopify's own Shop directory across 13 categories. Not a check for the presence of a file. A real cart on every single store, cancelled straight afterwards.

What "can an agent shop here" means

We ran the same test on every store, following the steps an agent actually takes:

1. Discover the store's agent surfaces: llms.txt, agents.md, the UCP profile, the catalog MCP endpoint.
2. Read the catalog the way an agent does, through the public product JSON and the UCP catalog tools.
3. Search for the store's own products using words taken from its own product titles.
4. Understand products: description, category, brand, images, SKU, barcode, and whether variants can be told apart.
5. Read policies: refund, shipping, privacy, terms, and whether the policy question tool answers.
6. Create a cart through the store's agent API with a real variant, update it, then cancel it.
7. Get a checkout URL the shopper could actually open.
8. Check the product pages for schema.org Product data with price and availability.

Every finding is labelled VERIFIED (we did it against the live store), OBSERVED (we read it but could not exercise it) or INFERRED. Only VERIFIED results are counted in the percentages below.

Two samples, one surprise

We expected the famous brands to lead. They trail on every row. The reason is not neglect, it is customisation. The mid-size merchants run Shopify as shipped, and Shopify ships agent-ready. The big brands run custom front ends, enterprise bot protection and bespoke domain setups, and every one of those decisions was made before agents existed.

The three ways stores fail

1. The agent endpoint is missing, or worse, broken

Thirteen of the 272 big brands had no usable agent surface, against one of 830 mid-size stores. They split into two groups, and the second is the interesting one.

Missing. At Ruggable, Huel, Vuori and LMNT, the UCP profile and the public catalog both answer 404 on the brand's own domain. We sent a shopping agent to Huel asking for a vanilla meal-replacement bundle under $100. It searched eight different ways and got zero products every time. It could not even confirm the store sells shakes:

"I searched this store eight different ways and every search returned zero products. I could not verify whether the store sells meal replacement shake powder at all."

The same thing happened at Ruggable, asked for a washable 8x10 rug that could handle a shedding dog.

Broken. Eight of the thirteen do have an endpoint, and it answers. It just answers with something an agent must refuse. At Ledger and Buck Mason the UCP profile returns a redirect, and the specification requires clients not to follow redirects at that address, so a compliant agent stops. At one cookware brand the profile returns a 200 with data that is not valid JSON. From the merchant's side everything looks fine: the URL resolves in a browser. From the agent's side the store may as well not be there.

In fairness to all of them: a custom storefront is a legitimate engineering choice, and these sites may serve human shoppers beautifully. The finding is narrow. On the domain a shopper would name, the agent surfaces that Shopify provides are not usable.

2. Bot protection turns away agents that identify themselves

Forty-five brands, 10% of those attempted, against six of 838 mid-size merchants. Our scanner announces itself honestly in its User-Agent as a read-only readiness test, and these stores answered 403 or 429 before we could read a single product. Whatever an agent platform sends, these merchants need an allowlist decision they have not made yet, and right now the default answer is no.

3. The cart works but the data does not

This is the largest group in both samples, and the one where the fixes are cheap. The agent layer is live, but:

• Product structured data with price and availability was present on 17% of brand stores and 21% of mid-size stores. Agents that browse pages rather than call APIs depend on it.
• Policy pages: only 36% of brands and 47% of mid-size stores had refund, shipping, privacy and terms pages an agent could find. Agents will not recommend a purchase when they cannot state the return terms. Across our simulated shopping trips the single most common gap was "the policy tool returned no shipping time".
• Descriptions: adequate on sampled products in 29% and 42% of stores. At Kylie Cosmetics, none of the 100 products we sampled carried a usable description. When our agent went looking there for a warm, cosy travel fragrance, it could not tell what most of the range smells like, and said so.
• Barcodes are not exposed publicly by any store in either sample.

Dollar Shave Club shows how thin the margin is. Its agent API returned a shave kit, in stock, at $27.99, then refused the cart with "Item cannot be purchased". The agent had a price and an availability flag it could not act on, so it stopped short of recommending the purchase.

What a good store looks like

Allbirds, asked for "a waterproof women's shoe in black for rainy commutes, size 8, under $170": the agent found six candidates, picked the Wool Runner NZ Mid Waterproof at $160, confirmed size 8 was in stock, created the cart and reached checkout, with high confidence. The whole run cost 16 cents.

"The Women's Wool Runner NZ Mid Waterproof in Natural Black is the best match: $160, described by the store as fully waterproof, and size 8 is in stock."

Gymshark, asked for black flare leggings in a short length under $60, went the same way, and so did Fashion Nova with a black going-out dress under $50. The best scores we recorded were Knockaround at 31 of 32 tasks among the brands, and Reigning Champ, Nation LA and Muscle Nation at 31 of 32 among the directory merchants.

The pattern across all twelve simulated shopping trips is consistent. When the agent layer and the product data both work, the agent commits to a recommendation with a real cart behind it. When either is missing, it hedges, tells the shopper to go and check the website, or gives up.

By category

From the Shop directory sample. Categories with fewer than 20 stores are left out of this table but counted in the totals.

Women's fashion and electronics lag because they are where custom themes and third-party checkout apps concentrate. Beauty stores pass the cart step but score lowest on tasks overall, almost entirely on descriptions and structured data: shade names with no words behind them.

What to do this week if you run a Shopify store

1. Scan your store. It is free, takes about a minute, and creates then cancels one test cart.
2. Check your agent endpoints on the exact domain a shopper would type, and check what they return, not just that they resolve. A redirect or a page of HTML where an agent expects JSON is the same as nothing at all.
3. Publish the four policy pages with concrete terms, and make sure your policy tool can answer a shipping-time question.
4. Fill in descriptions, product type and image alt text. These are the exact fields agents quote back to shoppers.
5. Decide deliberately which agents your bot protection admits, rather than letting the default decide for you.

Method and data

Samples. Brands: 462 well-known direct-to-consumer names we believed to be on Shopify; 272 were confirmed Shopify and scanned, 45 blocked us, and the rest turned out not to be on Shopify, to be unreachable, or to be closed. Directory: 838 merchants collected from Shopify's Shop directory across 13 categories on 24 September 2026, of which 830 were scanned.

Scanner. AgentShoppable v0.1, 39 checks, public storefront mode, no store credentials of any kind. Each storefront is resolved first, so we scan the exact host a shopper lands on, including www, and stores that are password-locked or closed are excluded rather than counted as failures.

Repeat scans. Where a store was scanned more than once and the results disagree, we keep the more complete run. A partial result indicates transient rate limiting, not a store that lost capability in half an hour.

Simulator. Claude Opus 5 acting as a shopper with read-only tools (search, get product, ask policy), followed by a deterministic cart test of whatever it picked. Twelve runs, total model cost $1.97.

Ethics. One cart per store, cancelled immediately. No orders, no accounts, no checkout sessions opened. Rate-limited throughout, and any store that blocked us was left alone. Every request that tests a store carries our own User-Agent, naming the tool and linking here. The one exception is the redirect-resolution request that works out which host to scan, which used a browser User-Agent for this dataset; the scanner now uses our own for that step too.

Caveats. Agent surfaces change daily, so these are point-in-time results. "Well-known" is our judgement, not a revenue ranking. Stores that blocked us may work perfectly for agents they have allowlisted. The directory sample skews toward merchants who opted into Shop, which probably flatters it a little.

Questions, corrections or a store you think we got wrong: hello@agentshoppable.com. If your store appears here and has since been fixed, tell us and we will rescan it and update the page.