Audit any ecommerce store for AI shopping-agent readiness, from the command line.
Shopping agents (ChatGPT, Claude, Perplexity, and the checkout copilots built on
top of them) do not browse stores the way humans do. They read robots.txt,
look for machine-readable catalogs, parse schema.org markup, and need a
deterministic way to hand a cart back to a human for payment. Most stores were
never built for that. This tool measures how ready a store actually is.
| check | weight | what it looks at |
|---|---|---|
products.json |
20% | Is there a structured, machine-readable catalog? For Shopify stores: coverage of descriptions, images, prices, SKUs, availability flags across products and variants. |
| structured data | 16% | schema.org JSON-LD on the homepage and a sampled product page (Product / ProductGroup, price, organization identity). |
| agent access | 14% | Fetches the homepage and a sampled product page as ChatGPT-User, Claude-User and OAI-SearchBot identify themselves, and compares each with a normal browser fetch. Flags blocks (403 and friends), bot-challenge pages, and near-empty shells. Robots rules say what a store intends; this shows what its CDN actually does. |
robots.txt |
12% | Blocked AI bots, weighted by what the block costs: shopper-driven fetchers (ChatGPT-User, Claude-User, Perplexity-User) cost the most, answer-engine indexers (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Amazonbot) cost less, and model-training crawlers (GPTBot, ClaudeBot, Google-Extended, CCBot, ...) barely count - opting out of training is a fair choice for a store. Declared sitemaps earn 10 points. |
| checkout handoff | 12% | Can an agent read cart state (/cart.js) and deep-link a variant into a cart (/cart/{variant}:1) without scraping? Probed with GET deep links in a throwaway session - the probe creates a session-scoped cart (that session cart IS the handoff mechanism) but never POSTs, never completes checkout, and never touches store state any shopper can see. |
| page structure | 10% | Homepage parseability: single <h1>, image alt-text coverage, semantic landmarks, text-to-markup ratio. |
llms.txt |
7% | Whether /llms.txt exists as a real text file (a catch-all HTML page doesn't count), whether it has a title, a one-line summary and described links, and whether the homepage links to it (<a> or <link>) so agents don't have to guess. |
| status codes | 5% | One GET to a made-up URL: a missing page should say 404, not 200 or a redirect home. Also flags bot challenges served as HTTP 200 and 429s without Retry-After. |
| server-rendered facts | 4% | For non-Shopify or headless storefronts: are the product name and price in the HTML the server sends, or only after JavaScript runs? Skipped on Shopify stores with an open catalog feed. |
Each check scores 0-100; the total is the weighted sum, graded A-F. A skipped check drops out and the other weights scale up to fill its share.
The stores that win agent-driven demand will be the ones agents can actually use: a clean product feed beats a beautiful hero video when the buyer is an LLM. This is the audit I run before touching a store's agent-readiness work - the output is the prioritized fix list, not the score.
git clone https://github.com/DeepanshuPal/agent-ready.git
cd agent-ready
pip install -r requirements.txt
python3 agent_ready.py https://www.allbirds.com
python3 agent_ready.py https://www.gymshark.com --jsonRequires Python 3.9+ and requests. No other dependencies.
Audit a whole list of stores and get one ranked report:
# stores.txt: one store URL per line (# comments and blank lines ok)
python3 agent_ready.py --bulk stores.txt --out audits/run-1/results --workers 5Stores are audited in parallel; one broken or unreachable store is recorded as an error row and does not kill the batch. Output is a CSV (one row per store: rank, score, grade, per-check scores, top fix) plus a JSON report with the same rows and batch aggregates (median/mean score, audited vs failed counts). This is how the 50-store run was produced.
Against gymshark.com (full output in examples/):
agent-ready report: https://www.gymshark.com
platform: Shopify score: 31/100 (F) audited in 2.2s
------------------------------------------------------------------------
[PASS] robots.txt 100/100 No AI shopping agents blocked in robots.txt
[FAIL] llms.txt 0/100 No llms.txt - agents get no curated map of the store
[FAIL] products.json 0/100 products.json not available (HTTP 403)
[WARN] structured data 45/100 Some schema.org markup, but no Product type found on the sampled product page
[FAIL] checkout handoff 0/100 Checkout handoff is limited - agents may not be able to complete purchases
[FAIL] page structure 45/100 Homepage parseability: 33% alt coverage, 6 landmark tags, 1658 KB HTML
------------------------------------------------------------------------
top fixes, in order of leverage:
1. Expose a machine-readable catalog (Shopify products.json, a Google Merchant feed, or a /products API).
2. Add schema.org Product JSON-LD (name, price, currency, availability) to every product page.
3. Publish /llms.txt describing the store, catalog endpoints and policies in plain markdown.
And allbirds.com, for contrast - a store that is genuinely close to agent-ready:
agent-ready report: https://www.allbirds.com
platform: Shopify score: 96/100 (A)
[PASS] robots.txt 100/100 No AI shopping agents blocked in robots.txt
[PASS] llms.txt 100/100 llms.txt present (87 lines of agent-facing docs)
[PASS] products.json 100/100 Machine-readable catalog exposed: 250 products, 100% with descriptions, 100% of variants priced
[PASS] structured data 90/100 Product pages use schema.org ProductGroup (newer type; older agent parsers looking for Product may miss it)
[PASS] checkout handoff 100/100 Agent can read cart state and hand off to checkout via deep links
[PASS] page structure 90/100 Homepage parseability: 62% alt coverage, 5 landmark tags, 655 KB HTML
Interesting spread in practice: Allbirds and tentree both publish llms.txt
already; Gymshark 403s its own products.json for scripted clients, which
means most shopping agents can't see its catalog at all.
- 50 DTC stores, 2026-09-10 - full ranked results, headline numbers, and the findings worth reading (including one store whose llms.txt tries to prompt-inject AI readers).
- Google Merchant Center feed linting (when a store has no products.json)
bulk mode: audit a list of stores, emit one CSV/JSON reportshipped 2026-09-12- UCP/ACP checkout-protocol probes
- non-Shopify platform depth (Magento, WooCommerce, custom headless)
MIT