README.md

---
license: cc-by-4.0
pretty_name: "ARC: can AI shopping agents reach checkout at these stores?"
language:
- en
tags:
- agentic-commerce
- ai-agents
- e-commerce
- shopping-agents
- benchmark
- ucp
- mcp
- webmcp
- tabular
size_categories:
- 1K<n<10K
configs:
- config_name: store_answers
  data_files: latest/store_answers.parquet
  default: true
- config_name: stores
  data_files: latest/stores.parquet
- config_name: rails_checks
  data_files: latest/rails_checks.parquet
- config_name: adoption_history
  data_files: latest/adoption_history.parquet
- config_name: personas
  data_files: latest/personas.parquet
- config_name: protocol_matrix
  data_files: latest/protocol_matrix.parquet
---

# 3,335 online stores checked: can AI shopping agents reach checkout at them?

**Snapshot `2026-10-12.r3`: 3,335 stores, checked 6 Oct 2026 to 12 Oct 2026.** The check finished for 3,318 of them (17 did not respond in time and are marked `not_tested`). Percentages below are of those 3,318.

| | Stores | Share |
|---|---:|---:|
| **AI agents can get a checkout link** through the store's agent connection | 1,311 | **39.5%** |
| AI agents can search products through an agent connection | 1,488 | 44.8% |
| No agent connection found: no product search, no valid UCP profile, not blocked (agents must click through the site like a person) | 1,425 | 42.9% |
| Bot protection blocked ARC's agent check | 348 | 10.5% |
| Publishes a valid Universal Commerce Protocol (UCP) profile | 1,601 | 48.3% |
| Shopify stores with a checkout link agents can use | 1,266 of 1,561 | 81.1% |
| Stores not detected as Shopify with a checkout link agents can use | 45 of 1,757 | 2.6% |
| robots.txt blocks at least one AI crawler (of 2,994 readable robots.txt files) | 153 | 5.1% |
| robots.txt blocks AI assistants fetching pages for a user (same base) | 101 | 3.4% |

All rows come from ARC's **automated HTTP check**. 82 stores also have a **browser cart test** (`browser_cart_test`). A checkout link is not a completed order: ARC never pays or places an order.

A free, open, per-store answer to one question: **can an AI shopping agent find products and reach checkout at this store?** Published by [ARC](https://www.arcreport.ai) at [arcreport.ai/data](https://www.arcreport.ai/data) under CC-BY-4.0. One frozen canonical snapshot per week. Cite the version folder, not `latest/`. The US checkout citation is `canonicalUsCheckout` in `src/lib/dataset/canonical.ts`, which reads `latest/manifest.json` and that version's `store_answers.csv`. The process is in `docs/runbooks/canonical-snapshot.md`. Nothing there is automated.

ARC is a free, neutral, CC BY open dataset and benchmark of how well online stores work with AI shopping agents, with a free advisory precheck API and an MCP connector.

## Links

- Site: [https://www.arcreport.ai/?ref=huggingface](https://www.arcreport.ai/?ref=huggingface)
- Connector (MCP): [https://www.arcreport.ai/mcp?ref=huggingface](https://www.arcreport.ai/mcp?ref=huggingface)
- Free API, no key: `/api/ask` (https://www.arcreport.ai/api/ask) and the advisory precheck:

```
curl 'https://www.arcreport.ai/api/precheck?url=https://allbirds.com'
```

Precheck results are advisory: they say whether an agent is likely to reach checkout. ARC stops before payment.

Headline numbers for each snapshot are in `summary.json` and `manifest.json` (`headline`), and at the top of [arcreport.ai/data](https://www.arcreport.ai/data). `summary.json` is computed from `store_answers.csv`.

| Count | Meaning |
|---|---|
| attempted | Every row in `store_answers.csv`. One store ARC attempted. |
| completed | The check finished (`can_agents_get_checkout_link` is `yes` or `no`). Every percentage uses this denominator. |
| failed | The check did not finish (`not_tested`). |
| stores_checked | The same number as completed. The protocol matrix UCP and MCP denominator on the current snapshot is this same completed count. |

## Valid UCP profile, US stores

**Snapshot 2026-10-12.r3, US stores with completed checks, n=1,271**

Shopify: 660 of 693 (95.2%). Non-Shopify: 50 of 578 (8.7%). All US: 710 of 1,271 (55.9%). A valid /.well-known UCP profile was served, not that checkout works.

Counted from `store_answers.csv` in snapshot 2026-10-12.r3. `market` is United States. A completed check is `can_agents_get_checkout_link` of yes or no. Shopify is `platform` equal to Shopify. Non-Shopify is every other `platform` value, including Not detected and empty. The numerator is `ucp_profile_valid` yes.

**Start with `store_answers`**: one row per store, plain yes/no answers, a one-sentence verdict, the single most useful fix, and a link to the store's page on ARC with the details.

## Column dictionary: `store_answers`

| Column | What it means |
|---|---|
| `domain` | The store's website, without www. |
| `store_name` | Store name from ARC's curated list. Empty for stores added by the public. |
| `category` | What the store sells (ARC's label). Empty when ARC has not labelled it. |
| `platform` | Shop software detected from the homepage (for example Shopify). "Not detected" means no known fingerprint (often a custom site). Empty means the homepage could not be read. |
| `market` | Where the storefront sells (ARC's curated label, not measured). Empty when not labelled. |
| `checked_at` | Date (UTC) of the latest automated check. |
| `can_agents_find_products` | Could an AI agent search the store's products through a machine-readable agent connection and get a product back (the specific product, for the ~90 Index stores where ARC tests one)? yes / no / blocked (the store returned a block to ARC's automated HTTP check on checked_at; this is a response to ARC's checker, not a finding against the store) / not_tested (the check did not finish). "no" is about agent connections only: an agent may still click through the website like a person. |
| `can_agents_read_price_and_stock` | Does the product page ARC opened list price and stock in structured data agents can read (schema.org Product)? yes / no / not_tested (no product page could be opened). |
| `can_agents_get_checkout_link` | Did the agent connection hand back a checkout link for a product it returned? yes / no / not_tested. ARC stops there: it never pays or places an order, so "yes" means an agent can reach checkout, not that an order was placed. |
| `agent_catalog_type` | Which agent connection answered with a product search tool: Shopify UCP MCP, Shopify Storefront MCP, or none. |
| `ucp_profile_valid` | Does the store publish a valid Universal Commerce Protocol profile at /.well-known/ucp? yes / no / unknown (blocked or errored). A valid profile is stricter than `publishes_ucp_file`: it needs a dated version and a shopping service. |
| `publishes_ucp_file` | Does the store publish a UCP file at /.well-known/ucp? yes = HTTP 200 and JSON with a version plus services or capabilities. no = HTTP 404 or 410. unknown = HTTP 403, 429, 5xx, a timeout, or a response that was not JSON. Empty means this check is not recorded. Unknown and empty never count as no. This means the store publishes a UCP file. It does not mean an agent can reach checkout there. |
| `ucp_file_checked_at` | UTC date of the /.well-known/ucp file check. Empty when that check is not recorded. |
| `webmcp` | Does the homepage declare WebMCP tools for in-browser agents? yes / no / not_tested (checks before 6 Oct 2026). |
| `blocks_ai_crawlers` | Does robots.txt disallow at least one of the AI crawlers ARC tracks (answer, search or training bots) from the homepage or tested product? yes / no / unknown (robots.txt could not be read). A rule in a file, not an observed block. |
| `blocks_ai_answer_agents` | Same, but only for AI assistants fetching pages for a user (for example ChatGPT-User, Perplexity-User). yes / no / unknown. |
| `browser_cart_test` | Result of ARC's separate browser cart test, where it exists (about 90 stores): for example "2 of 5 added to cart". Empty means no browser test. |
| `verdict` | A plain answer to "can AI agents reach checkout at this store?", the same sentence as the store's ARC page. The strongest answer is "AI agents can reach checkout": ARC's agent got a checkout link. ARC always stops before paying, so it never claims more than reached checkout. |
| `top_fix` | The most useful change for the store, in plain words, the same as the first fix on its ARC page. Empty when nothing is needed. |
| `checked_by` | How the answer was measured: "automated HTTP check" or "automated HTTP check + browser cart test". |
| `evidence_url` | The store's page on ARC with the full details. |
| `tested_url` | The page the check started from: the homepage for most stores, one product page for Index stores (query string removed). |
| `source` | How the store got into ARC: index (weekly browser Index), directory (ARC's curated list), submitted (added by the public) or checked (one-off check). |
| `snapshot_date` | Date of this snapshot. |
| `method_version` | Version of ARC's method used. |
| `direct_agent_integration` | Public files an agent can read without an account: detected, not_found, or unknown. Empty means not recorded. Unknown and empty are not a no. A detected file is not a claim that an agent can reach checkout. |
| `direct_agent_integration_checked_at` | UTC date of that file check. Empty when not recorded. |
| `guest_checkout_offered` | During the no-payment checkout check: yes, no, or unknown. Empty means not recorded. ARC does not create an account. Unknown and empty are not a no. |
| `account_creation_required` | Whether that page required an account and did not offer guest checkout. yes / no / unknown. Empty means not recorded. |
| `account_signup_without_captcha_or_phone` | Whether a signup form was already on the page with no captcha and no phone check. yes / no / unknown. ARC does not open or submit the form. |
| `account_guest_checked_at` | UTC date of the account and guest reading. Empty when not recorded. |
| `block_reason` | Recorded reason for a block on the automated HTTP check: robots_disallow, bot_wall, tls, or other. Empty when the check was not blocked, when no reason was recorded, or on a snapshot exported before this column. Empty is not a bot wall. Bot protection or a bot wall is named only when this cell is bot_wall. A block is a response to ARC's checker, not a finding against the store. |

## How it is measured

Two separate methods. Every row says which one it comes from (`checked_by`).

1. **Automated HTTP check (every store).** A plain program, not an AI model and not a browser, requests the store's machine-readable shopping interfaces the way an agent would: the Universal Commerce Protocol profile at `/.well-known/ucp`, Shopify's agent catalog endpoints (`/api/ucp/mcp`, `/api/mcp`), a product search, and the cart step that returns a checkout link. It also reads the homepage (platform fingerprint, WebMCP tags) and robots.txt. It identifies itself, runs at most about once per host per week, and never pays or places an order.
2. **Browser cart test (fewer than 100 Index stores only).** Five ARC browser test shoppers each try to put one real product in the cart through the normal website, stopping before payment. Shown in `browser_cart_test`. These are ARC's own test runs, not sessions of any named consumer AI product.

## Limitations (please read before citing)

- **Cite a figure with its n, its denominator, and its snapshot.** Field names, numerators, and what each headline count does not prove: [metrics glossary](https://www.arcreport.ai/methodology/metrics).
- **A checkout link is not a completed order.** ARC stops at the link. "yes" means an agent could reach checkout through the store's agent connection on the check date, not that an order would succeed.
- **"no" is about agent connections only.** A store without one can still be used by an agent that clicks through the website like a person. Our browser cart test covers that, but only for fewer than 100 stores.
- **One check, one moment.** Each answer is the store's latest check (dates in `checked_at`). Stores and stock change; a single failed check can be temporary.
- **"blocked" means the store returned a block to ARC's automated HTTP check on checked_at.** This is a response to ARC's checker, not a finding against the store. Real agents may or may not be treated the same way. Bot protection or a bot wall is named only when `block_reason` is `bot_wall`.
- **robots.txt is a rule in a file, not an observed block.** Bots may ignore it, and firewalls may block bots that robots.txt allows.
- **The store list is a sample, not all of e-commerce.** It is ARC's curated directory plus stores the public submitted. It leans toward English-language, direct-to-consumer brands and toward Shopify.
- **Labels such as `category` and `market` are ARC's curated labels**, not measurements.
- `not_tested` and `unknown` are used only where there is genuinely no result; they never mean "no".

## Other tables

The full technical detail stays available alongside `store_answers`:

| Table | One row per | What it holds |
|---|---|---|
| `rails_checks` | Store | The raw fields of the automated HTTP check (UCP profile status, MCP endpoints, catalog tool, cart status, WebMCP, platform id, evidence label) |
| `stores` | Index store | Browser cart test results: score, test shoppers that added to cart, failure label, per-persona outcome |
| `adoption_history` | Week × source × metric | Weekly counts |
| `protocol_matrix` | Protocol × stage | Declared / Valid / Usable / Transacts counts per protocol |
| `personas` | Browser test shopper | The five browser test personas |

Field reference for the technical tables is below. Method and changelog: [arcreport.ai/methodology](https://www.arcreport.ai/methodology).

## What is excluded

No scraped page text or product copy, no model-written prose, no costs, no emails, accounts or visitor data. Only ARC's own answers, signals, URLs and dates. OpenStreetMap-derived directory rows are left out: they are ODbL share-alike, and this dataset stays CC-BY-4.0.

Older snapshots in /data/history can include OpenStreetMap-derived hosts. Those hosts are © OpenStreetMap contributors under the Open Database License (ODbL), not CC-BY-4.0. Current CC-BY exports leave those hosts out.

Hugging Face card (abab7575/arc-open `docs/dataset-card.md`, same paragraph): OpenStreetMap-derived directory rows are left out of this CC-BY-4.0 dataset because they are ODbL share-alike. Older snapshots in /data/history can include those hosts. Those hosts are © OpenStreetMap contributors under the Open Database License (ODbL), not CC-BY-4.0.

## License & attribution

The data ARC publishes is licensed under the Creative Commons Attribution 4.0 International Public License ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)).

- Cite as: **ARC (arcreport.ai)**. Link: [https://www.arcreport.ai](https://www.arcreport.ai) (dataset: [https://www.arcreport.ai/data](https://www.arcreport.ai/data)).
- Terms: [https://www.arcreport.ai/terms](https://www.arcreport.ai/terms).
- Snapshot `2026-10-12.r2` includes 203 OpenStreetMap-derived rows under the [ODbL-1.0](https://opendatacommons.org/licenses/odbl/1-0/). Attribution: © OpenStreetMap contributors. Later snapshots leave those rows out.
- Third-party data is not relicensed. Tranco ranks and lists are not published and are not licensed by ARC.
- Wikidata-derived fields are CC0.
- Host names discovered via Common Crawl: source cited as Common Crawl (commoncrawl.org).
- Unpublished discovery-source data is excluded. Hosts that opted out are excluded.
- CC BY 4.0 covers only ARC's own measurements.
- Results are advisory. ARC stops before payment.
- Code is MIT (github.com/abab7575/arc-open).

Contact: arcreport@mail.grokbot.com. Store owners can ask for a correction or removal at [arcreport.ai/contact](https://www.arcreport.ai/contact).

## Files and versioning

Each snapshot lives in its own immutable version folder and is copied to `latest/`. Every table comes as CSV and Parquet, with a `manifest.json` holding row counts, headline numbers, `snapshot_version`, `generated_at` (UTC) and `supersedes`.

### Versioning

- A version folder never changes once published. The first export on a date is `YYYY-MM-DD/`; any later export on the same date is published as `YYYY-MM-DD.r2/`, `YYYY-MM-DD.r3/` and so on. The exporter refuses to overwrite an existing version.
- `latest/` is a moving copy of the newest version. To cite exact numbers, cite the version folder (for example `/data/2026-10-07.r5/`), not `latest/`. On 13 Oct 2026 the answer text was corrected in `latest/store_answers` and in the 2026-10-09 and 2026-10-12 folders. `2026-10-09` has no OpenStreetMap-derived rows. `2026-10-12.r2` is mixed licence (CC BY 4.0, except 203 OpenStreetMap-derived rows under ODbL-1.0, attributed to OpenStreetMap contributors). `2026-10-12.r3` drops those rows and is the CC BY 4.0 edition `latest/` cites. Neither date's folders contain unpublished discovery-source rows. See the erratum in `/data/history/README.md`.
- `versions.json` lists every version with its `generated_at`, `supersedes`, method version, stores checked and row counts.
- Before this rule (7 Oct 2026), `/data/2026-10-07/` was regenerated four times during the day (rails_checks went from 607 to 647 to 1,044 rows). All five published versions are restored byte-for-byte from the commits that published them, as `2026-10-07/` (r1) to `2026-10-07.r5/`; their manifests carry `restored.source_commit`.

## Technical tables: field reference

The technical tables carry `snapshot_date`, `method_version` and `evidence` (`measured` or `inferred`).

### `stores`
| Field | Meaning |
|---|---|
| slug, store_name, domain, category | Store identity from ARC's public Index list |
| homepage_url, product_url | The store homepage and the one product the checkout check starts from (query strings removed) |
| score | ARC compatibility score, 0–100. Empty when a retest is pending |
| agents_chose, agents_total | How many ARC browser personas added the product to the cart, out of how many |
| failure_reason | Fixed label for where agents stopped (e.g. `Blocked at the size picker`), `Added to cart (stopped before payment)` when every shopper got the product into the cart (ARC never pays; older snapshots said `Finished the purchase`), or `other` |
| persona_a … persona_e | Per-persona outcome: `chose`, `missed` or `error`. These are ARC browser shopper personas run on ARC-chosen models, not named consumer agents. Legacy alias ids (muse, instinct, grok, openai-dots, openclaw) are in `personas.csv` (`legacy_alias_id`) |
| index_week, scanned_at | ISO week and UTC time of the scan |
| evidence | `measured` when all five shoppers ran and a score exists; `inferred` otherwise |

### `rails_checks`
| Field | Meaning |
|---|---|
| domain, slug, category, product_url | Store checked and the product page used (query strings removed; cart URLs are never stored) |
| status | `completed` or `failed` |
| webmcp, storefront_mcp, ucp_mcp | Whether each interface was detected. `webmcp` is empty (not measured) for checks before 18:00 UTC on 6 Oct 2026, when detection started |
| catalog_tool, catalog_found | Catalog tool exposed and whether it returned the product |
| cart_status | `link`, `unavailable`, `not_offered`, `error` or `skipped` |
| ucp_profile | `/.well-known/ucp` validity: `valid`, `invalid`, `absent`, `blocked` or `error`. Stricter than `ucp_file`. |
| ucp_file | `/.well-known/ucp` file: `detected` (HTTP 200 JSON with a version plus services or capabilities), `not_found` (404 or 410), `unknown` (403, 429, 5xx, timeout or non-JSON), or empty when not recorded. Empty and unknown are not a no. This is a published file, not proof an agent can reach checkout. |
| mcp_blocked | The store blocked the MCP request |
| checked_at | UTC time of the check |
| evidence | `measured` for a clean completed check; `inferred` when blocked, errored or failed |
| source | `index` (ARC Index store), `directory` (curated list), `submitted` (added by the public via /agent-checkout/submit, the API or MCP, after validation) or `checked` (one-off check only) |

### `adoption_history`
`week` (ISO), `source` (`index` or `rails`), `metric` (e.g. `stores_scored`, `mean_score`, `stores_any_persona_carted`, `ucp_profile_valid`, `ucp_file_detected`, `ucp_file_not_found`, `ucp_file_unknown`, `cart_link`, `webmcp`, `webmcp_measured`) and `value`. The UCP file adoption count is `ucp_file_detected` out of `ucp_file_detected` plus `ucp_file_not_found`. `ucp_file_unknown` is separate. A missing result is not counted.

One definition with `stores`: index metrics use the latest completed scan per published Index store within the week, so the current week matches `stores` exactly. Rails metrics count each host once per week (its latest attempt that week). `webmcp` counts only checks after detection began; its denominator is `webmcp_measured`.

### `personas`
`persona`, `label`, `style`, `legacy_alias_id`, `method_version`.

### `protocol_matrix`
See [arcreport.ai/protocols](https://www.arcreport.ai/protocols).

## Host discovery sources

Checker host lists may include hostnames found in the Common Crawl columnar URL index (`cc-index:` plus the crawl id). Data from Common Crawl (https://commoncrawl.org). Hostnames only. Common Crawl page content is not stored or republished. Common Crawl terms are at https://commoncrawl.org/terms-of-use. Exported source tags are limited to `cc-index:`, `wikidata:`, `osm:`, `expansion:`, and `source:wikidata`. Discovery-source, campaign, and `source:tranco` tags are not exported. A stored `source:tranco` tag stays on the row.

## Files in this snapshot

Mirror of the official ARC open dataset at [arcreport.ai/data](https://www.arcreport.ai/data), updated after each weekly snapshot (Mondays). Files are byte-for-byte copies of ARC's published CSV and Parquet.

**Current snapshot:** `2026-10-12.r3` (snapshot date 2026-10-12, generated 2026-10-13 19:30 UTC, method `arc-method-2026.10`).

| Config / table | Rows | What it holds |
|---|---|---|
| `store_answers` | 3,335 | Start here. One row per store in plain words: can AI agents find products and get a checkout link, what blocks them, a one-sentence verdict, top fix, and whether the answer is from the automated HTTP check or also the browser cart test. |
| `stores` | 89 | Latest completed ARC Index result per store (browser cart test: score, personas that carted, failure label). |
| `rails_checks` | 3,335 | Latest HTTP agent check per store domain (UCP profile, WebMCP, storefront MCP, catalog tool, cart link, checkout WebMCP). |
| `adoption_history` | 32 | Weekly aggregate counts by source and metric (long format). |
| `personas` | 5 | The five ARC browser shopper personas (A to E) with legacy alias ids. |
| `protocol_matrix` | 24 | Declared / Valid / Usable / Transacts counts per protocol with evidence labels. |

```python
from datasets import load_dataset
answers = load_dataset("ArcReport/arc-agent-commerce", "store_answers", split="train")  # start here
```

Or just download [`latest/store_answers.csv`](https://huggingface.co/datasets/ArcReport/arc-agent-commerce/resolve/main/latest/store_answers.csv).

### Versions
- `latest/` is the newest snapshot. **To cite exact numbers, cite a version folder** under `versions/` or the Hub tag `v<version>` (e.g. `v2026-10-12.r3`), not `latest/`.
- Version folders never change once published. A second export on the same date is `YYYY-MM-DD.r2`, `.r3` and so on. `versions.json` lists every version.

- `2026-10-07` (generated 2026-10-07 08:08 UTC, method arc-method-2026.10)
- `2026-10-07.r2` (generated 2026-10-07 11:27 UTC, method arc-method-2026.10)
- `2026-10-07.r3` (generated 2026-10-07 11:41 UTC, method arc-method-2026.10)
- `2026-10-07.r4` (generated 2026-10-07 14:32 UTC, method arc-method-2026.10)
- `2026-10-07.r5` (generated 2026-10-07 19:17 UTC, method arc-method-2026.10)
- `2026-10-09` (generated 2026-10-09 15:47 UTC, method arc-method-2026.10)
- `2026-10-09.r2` (generated 2026-10-09 15:54 UTC, method arc-method-2026.10)
- `2026-10-09.r3` (generated 2026-10-09 16:03 UTC, method arc-method-2026.10)
- `2026-10-12` (generated 2026-10-12 13:34 UTC, method arc-method-2026.10)
- `2026-10-12.r2` (generated 2026-10-12 20:37 UTC, method arc-method-2026.10)
- `2026-10-12.r3` (generated 2026-10-13 19:30 UTC, method arc-method-2026.10)

### Citation

```bibtex
@misc{arc_agent_commerce_index_2026,
  title  = {ARC: can AI shopping agents reach checkout at these stores? Open dataset},
  author = {{ARC}},
  year   = {2026},
  note   = {Snapshot 2026-10-12.r3, method arc-method-2026.10},
  howpublished = {\url{https://huggingface.co/datasets/ArcReport/arc-agent-commerce}},
  url    = {https://www.arcreport.ai/data},
  license = {CC-BY-4.0}
}
```