How ARC tests whether agents can buy from your store

ARC tests agent buyability in two lanes. First, a live HTTP check asks whether your store's catalog is readable by agents. Second, five ARC browser cart tests try to add a product to the cart and move toward checkout, recording the exact step where each one stops.

On top of those two lanes, ARC builds a Reachability Matrix for named agents, checks what an AI model tells shoppers about you, and runs early detection for Personal Agent Protocol markers. This post explains each piece and where the method falls short.

If you're new to the term, start with What is agent buyability?

Lane 1: the live HTTP agent-catalog check

Many agents don't load your pages like a person. They send plain HTTP requests and read what comes back.

ARC makes live HTTP requests to your store the way an agent would. It:

• loads your homepage (and can detect Shopify WebMCP in the HTML)
• looks for a storefront MCP at /api/mcp and a UCP MCP at /api/ucp/mcp
• reads your UCP profile, if present
• calls catalog search tools (for example search_catalog or search_shop_catalog) and records whether your product was found
• may call cart tools (create_cart / update_cart) only — enough to get a checkout continue link

It never logs in, never opens a browser, never calls payment or order tools, and never places a purchase. A returned checkout link means an agent can reach checkout, not that a purchase completed. Results for a host are cached for 24 hours. If MCP or UCP is blocked, ARC records that too.

Lane 2: ARC browser cart tests

Some agents do drive a browser. They click through product pages like a shopper. So ARC runs that path too.

Each scan runs five ARC browser cart tests. Each one is an automated browser cart attempt with a shopper persona. It opens a product, picks options, adds to cart, and moves toward checkout. ARC Index tests stop at add-to-cart. Every test stops before entering payment details and never places a real order. They also stop on CAPTCHA, OTP/MFA, 3D Secure, or a login wall without credentials.

For every attempt, ARC records milestones in order: product → variant → cart → checkout → confirmation. The stuck step is the first failed milestone. In plain language, that often looks like:

• the product page never loads for automated traffic
• variant selection (size, colour, shade) can't be completed
• add-to-cart doesn't respond
• a pop-up or interstitial blocks the flow
• checkout can't be reached

Checkout handoff is not a completed purchase. You see which step failed and a specific fix for it.

In ARC's October 2026 Index snapshot, variant selection (the size or colour picker) was the published stop reason at 8 of 83 scored stores (10%). The add-to-cart step itself was more common, at 39 of 83. Index tests stop at add-to-cart (State of agent buyability: October 2026).

The two lanes often disagree. In the same snapshot, 63 of 83 scored stores returned an agent checkout link over HTTP. At 37 of those 63, at least one ARC browser cart shopper still failed to add the product to the cart. At 14 of them, all five failed: the agent rails said yes and the storefront said no.

Free scan, no login: arcreport.ai. Both lanes run on your homepage in about a minute.

The Reachability Matrix

Your store hub at /store/<your-domain> includes a Reachability Matrix. Rows are named agents:

• Muse
• Instinct
• ChatGPT
• Google AI Mode
• Grok Bot
• OpenClaw

Columns are three questions:

• Discover: can this agent find your store and products?
• Buy: can it get through to checkout?
• Allow/block: do your site's rules let it in, or is it at risk of being blocked?

Every cell carries an evidence chip:

• Measured: ARC ran a direct check — an HTTP rail check or an Index browser cart test. This is not a live Muse, Instinct, ChatGPT, Gemini, Grok, or OpenClaw shopping session.
• Inferred: ARC reasoned it from other signals, not a direct test of that named agent. Example: a valid UCP profile can make Google AI Mode Discover and Buy cells Inferred ("may enable… for eligible merchants; no live Gemini test").
• Unknown: no evidence yet. The UI labels this "Not confirmed."
• Public report: a public report is available for that cell.

The chips matter. They tell you how much weight to put on each cell. We would rather show you Inferred or Unknown than dress up a guess as a fact.

Real model read-back

Reach and checkout are only part of it. The agent also has to get your facts right.

ARC makes a real API call to Grok and asks what it would tell a shopper about your store. ARC then checks that answer against your own pages on five fields only: price, returns window, shipping cost, availability, and delivery estimate. Evidence prefers JSON-LD, Shopify product JSON, or the buy box for price and stock, and policy pages for shipping and returns. Each field comes back as match, mismatch, grok didn't know, or couldn't verify.

It shows you where your public information is unclear or easy to misread. Product name is not a verified field.

The PAP lane: early detection only

On 6 October 2026, Sierra and Meta announced the Personal Agent Protocol, an open standard for how personal agents interact with businesses. Sierra plans to publish the v0.1 specification later in October 2026 (Sierra, "Introducing Personal Agent Protocol").

The spec isn't published yet. So ARC's PAP lane is early detection. It looks for discovery markers a store might expose ahead of v0.1 and reports what it sees.

It makes no support claim. ARC does not certify stores for PAP. When v0.1 lands, we'll update this lane.

What ARC does not do

Plainly:

• ARC does not run live Muse or Instinct sessions on every store. We run spot checks only. The default test on every store is the HTTP catalog lane plus ARC browser cart tests.
• ARC is not affiliated with any agent vendor. That includes Meta, Sierra, Instinct, OpenAI, Google, and Shopify. Agent names in the matrix describe which agents we assess, not partnerships.
• ARC does not track AI mentions or share of voice. Mention tools do that well. ARC starts where they stop: can the agent buy?
• ARC does not certify protocol support. See the PAP lane above.

Limits of this method

Honesty about limits is part of the product. The full method is on the methodology page. Here is what to keep in mind:

• ARC browser cart tests are not real agents. They follow the same kind of path a browser agent would. A specific agent may behave differently, better or worse.
• Inferred cells are reasoned, not observed. Treat them as a strong hint, not proof.
• Five attempts is a sample. It catches common, repeatable failures. It can miss rare ones or problems on products we didn't test.
• Results are a snapshot. Sites change. A theme update can break checkout tomorrow. Each result that has evidence shows its test date (empty new hubs with no rails or Index data yet won't).
• Read-back covers one model. Grok's answer is a useful check, not a stand-in for every AI model. Re-tests do not re-run Grok.

ARC is a free, open research project. Every scan, page and dataset is free. The HTTP catalog check is cached for 24 hours and does not change the score.

---

See what agents hit on your store. Free scan, no login: arcreport.ai. Then see your store hub: arcreport.ai/store.

ARC is independent and not affiliated with Meta, Sierra, or any agent vendor.