AgentBI

Is your website reachable, readable, and purchasable by an AI agent?

• Web: https://agentbi.tech — paste a URL, get a report
• CLI: agentbi https://your-store.com

Agentic commerce protocols (UCP / ACP / AP2) shipped in 2025–2026. This answers a much more boring question: how many sites have actually deployed any of it, and what does a single site need to fix.

What it checks

Output is a 0–100 score plus a prioritized fix list.

Scoring is heuristic. It is for seeing the gap, not for auditing a site.

Install

pip install agentbi           # or: uvx agentbi <url>  (nothing installed permanently)

agentbi https://your-store.com          # human-readable report
agentbi https://your-store.com --json   # machine-readable
agentbi https://your-store.com --no-probe   # static checks only, no multi-UA probing

From a clone (no install): python3 agentbi.py https://your-store.com — same flags.

Web version (stdlib only):

python3 serve.py --port 8000     # serves web/ + /api/check + /leaderboard

Limitations — read before quoting any number

These are the ones that matter; the leaderboard page says the same thing.

1. No JS rendering. Client-rendered content is invisible to this scanner.
2. It does not place an order. A UCP manifest is checked for existence and version, not for actually completing a checkout.
3. Coverage is ~50%, and it is our problem, not the sites'. In the 2026-10-05 batch, 115 of 225 stores were evaluable; 78 returned 403. That is datacenter IP reputation hitting enterprise WAFs. Full browser headers were tried and did not reproduce. The data covers stores without an enterprise WAF, not the industry.
4. WooCommerce n=8. It is every site harvestable from WooCommerce's own showcase. Read it as a direction, not a percentage.
5. llms.txt is often platform-generated. Allbirds, Death Wish Coffee and Ombraz have byte-identical templates. On Shopify it is a platform dividing line, not a signal of effort.

Data

batch.json — 225 online stores scanned 2026-10-05 (115 evaluable), with per-site platform, score, UCP version, JSON-LD and llms.txt findings. Rendered at https://agentbi.tech/leaderboard.

Headline: Shopify 69 stores → 59 with UCP (86%), avg 78. WooCommerce 8 → 0, avg 28. Custom stacks 35 → 1 (avg 14). Two stores were one UCP version behind.

Language

English by default (the audience is store owners, mostly non-Chinese). ?lang=zh on the site, --lang zh on the CLI. All user-facing strings live in agentbi_strings.py — add keys in both languages, a missing key silently falls back to English.

Crawling behaviour

• Content is fetched only when someone submits a URL. No continuous crawling.
• User agent is honest and identifiable: agentbi/0.1 (+https://agentbi.tech/about; agent-readiness scanner)
• The single exception is the multi-UA probe, which must use real agent identifiers to measure differential treatment. That step is documented on /about.
• ~10 requests per check, 15–20s. One real scan per domain per 10 minutes; results cached 30 min.
• Want out? contact@agentbi.tech.

License

MIT, except batch.json data which you are free to use with attribution.