How ARC measures agent buyability

Two lanes

ARC tests one question two ways: can a personal agent buy from this store? It measures how far an agent gets, up to a checkout link or a filled cart, and always stops before paying. The browser laneis ARC's own browser cart test. The agent-catalog laneis a real HTTP check of the store's agent catalog. The published score is the browser lane. The catalog check never changes the score. The verdict uses both lanes: agents reach checkout when the catalog returned a checkout link or a browser cart shopper added the product to the cart.

The browser cart test

For each store on the ARC Index, ARC opens one named product page in a real headless browser. Five ARC browser cart shoppers then try to add that product to the cart: ARC browser shoppers A to E. They are ARC personas, not the consumer agents of the same style. Older data used product-style ids (muse, instinct, grok, openai-dots, openclaw); those survive only as a documented alias in the dataset.

• Each shopper is an ARC persona with its own instructions (for example, social and visual, price-driven, or a scripted crawler that struggles with JavaScript-heavy pages). On Index and free scans all five personas run on the same low-cost model, with a short step and time budget.
• This is nota live session of any consumer shopping agent. No one's account or app is used.
• Shoppers behave like a careful visitor: they pick the first in-stock size or colour, dismiss cookie and country pop-ups, stay on the store's own site, never log in, never attempt a CAPTCHA or one-time code, and stop before any payment details.
• A shopper "chose you" when the product verifiably landed in the cart. Nothing is bought. "5 of 5" means all five shoppers added the product to the cart.
• When shoppers stop, ARC classifies where (for example, the size picker, a cookie banner, a human-check, the add-to-cart button, or a block at the door) and publishes it with HTTP or screenshot evidence.
• If ARC cannot stand behind a result (for example, a problem on ARC's side), the store shows Retest pending. It is not ranked, and a missing score is not a zero.

How the score is computed

Each shopper's journey gets a depth from 0 to 5: blocked at the door (0), reached the store (1), navigated (2), opened the product (3), added to cart (4), reached checkout (5). Then:

• Breadth (B)is the average depth across the shoppers, divided by 5. A shopper that failed because of an error on ARC's side is left out.
• Machine readability (M)is a 0 to 100 check of the product page's raw HTML: structured data such as JSON-LD, schema.org, and Open Graph, read without rendering the page.
• Score = 100 × (0.7 × B + 0.3 × M), then capped so completion dominates: if no shopper added to cart the score is scaled into 0 to 49; if fewer than half did, it is capped at 59; if fewer than three quarters did, at 74; if no shopper reached the store, at 25.
• Index tests end at add-to-cart, so the highest score a store can reach on the Index today is 86 (all five at add-to-cart, perfect machine readability).
• Public Index pages show the score out of 100 and how many of the five shoppers added the product to the cart. Older one-off reports also show a letter band (A is 90 and up, B 75 to 89, C 50 to 74, D 25 to 49, F under 25) as a reading aid on the same score; it is not a separate measure, and because Index tests stop at add-to-cart no Index store can reach an A.

The agent-catalog check

A real HTTP request, identified as ARCBot. ARC lists the tools on the store's UCP MCP endpoint (/api/ucp/mcp) and its storefront MCP endpoint (/api/mcp), searches the catalog for the tested product, and asks the cart tool for a checkout link, stopping before payment. Shopify moved catalog and cart tools from /api/mcp to UCP, so ARC uses the UCP tools first. It also reads the Universal Commerce Protocol profile at /.well-known/ucpand detects the Shopify WebMCP storefront adapter. An HTML page or a bot block is not counted as a profile. There is no model call, no order, and no score change. The checkout link itself is not stored. The check is Shopify-shaped, so "no agent catalog" on another platform does not prove agents cannot buy. Try it free at /agent-checkout. Protocol-by-protocol counts (Declared, Valid, Usable, Transacts) are on the Protocol Matrix; definition changes are logged in the method changelog.

Checkout WebMCP (2026-10-08)

Shopify stores receive a GET-only checkout HTML check using an available variant and isolated cookies, within the same per-store deadline. Checkout exposes WebMCP tools (get/update/complete, buyer approval required) when initialization markers appear; declared tool names are recorded. This is HTML detection, not a live browser session or proof that every tool is ready. ARC never executes JavaScript or checkout tools, supplies buyer data, or purchases anything. Only same-store redirects are followed; a shop.app redirect is read for its same-store ur_back_url. Checkout tokens and query strings are never saved. Rates include Shopify stores whose checkout was measured. Unknown failures and not-checked stores are separate from measured no; other platforms are not applicable. Completed checks missing evidence are backfilled after 24 hours, with never-checked stores first.

Agent Reachability Matrix

Every store hub shows a matrix with one row per ARC browser shopper (A to E) and one row for the Universal Commerce Protocol path (for example Google AI Mode / Gemini), across three columns: Discover, Buy / checkout, and Allow vs block risk. Rows are ARC personas, not named consumer agents; their legacy alias ids only keep old links working. Each cell carries an evidence chip:

• Measured: observed by an ARC check, either the HTTP catalog check (shared discovery rails, a returned checkout link, a blocked endpoint) or the published browser cart test for that shopper (added to cart, or not).
• Inferred: used for the UCP path. A valid Universal Commerce Protocol profile may enable discovery and checkout on UCP surfaces such as Google AI Mode for eligible merchants. ARC does not run a live Gemini test.
• Unknown: no evidence either way. No block evidence does not mean a store allows agents.
• Public report: a block of a named consumer agent reported in the press, not measured by ARC, shown in its own row with the source linked.

A detected rail is not a promise that any consumer agent can use it, and no cell is a live session of a consumer agent.

Personal Agent Protocol watch

The Personal Agent Protocol v0.1 specification is not published yet, so ARC does not claim PAP support for any store. Each store hub checks the homepage for PAP-shaped links or meta tags and requests a short list of candidate well-known paths (such as /.well-known/pap and /.well-known/personal-agent). Soft-404 HTML pages are ignored. With no match the status is "Watching for PAP v0.1". A match is shown as a possible discovery marker, a watch signal only.

One model's read-back

Where shown, this is one AI model's read-back (currently Grok via API), checked against the store's own pages. It is one model, not a consensus, it is not the consumer Grok app, and it does not change the score. Headlines describe it as one model's read-back.

Dataset definitions

• Scored store: a published Index store whose latest completed scan has a score. Any persona carted: at least one of the five browser personas got the product into the cart in that scan. stores.csv and theindex rows of adoption_history.csv use this same definition (latest completed scan per store, per ISO week), so the current week always matches.
• Rails rates count each host once, using its latest check. WebMCP is counted only for checks after detection began (18:00 UTC, 6 Oct 2026).
• Every change to these definitions is logged in the method changelog before numbers built on it are published.

Limits

• Each Index store is tested on one product. A result is a snapshot of that product on that day.
• The Index is a curated list of direct-to-consumer brands, weighted toward apparel and toward Shopify stores.
• ARC is independent. Not affiliated with or endorsed by Meta, Spear Street Technology, SpaceXAI, OpenAI, or the OpenClaw Foundation. ARC is also not affiliated with Sierra, Shopify, or Google.

Reports that use these numbers are published on the ARC blog. Method version arc-method-2026.10; see the method changelog.