Period: 2026-09 · Sample: 206 API companies · Method: reproducible, single command
An AI assistant's ability to buy from a company is now measurable. This report measures it, on the same corpus every month, and publishes the number even when it is unflattering to us.
Reproduce any row: node audit.js <domain> · Free live check: http://agentcheck.91.99.165.95.sslip.io
| Finding | Result |
|---|---|
| Companies that can accept payment from a machine (a live HTTP 402 challenge on their API) | 0 of 206 |
| Companies that finish every gate in the funnel | 0 of 206 |
| Companies publishing any agent surface (agent card, MCP descriptor, payment manifest) | 12 of 206 |
| Companies scoring zero on "can a machine pay you" | 129 of 206 |
| Highest score in the sample | 85/100 |
| Median score | 55/100 |
Companies with no llms.txt | 93 of 206 |
| Companies blocking named AI agents in robots.txt | 5 of 206 |
| Sites that refused an agent-style request outright | 17 of 206 |
The finding that matters: machine discovery is half-solved and cheap — almost everyone publishes a sitemap, and most publish docs a model can read. Machine payment is untouched. Not one company in the sample completes the last step, which is the only step that produces revenue.
| Band | Companies |
|---|---|
| 80–100 (near-ready) | 8 |
| 60–79 | 78 |
| 40–59 | 89 |
| 0–39 | 31 |
Nobody is in the top band. The ceiling is not engineering talent; it is the missing payment layer.
Every failing check across all 206 audits, most common first, with the artifact that fixes it:
| Failing check | Companies | Fixed by |
|---|---|---|
| Purchasable · Purchase possible with no human signup | 189 | x402-server.js (402 pay-per-call) |
| Payable · Machine-native payment challenge (HTTP 402) | 189 | x402-server.js |
| Payable · Agent surfaces published | 177 | .well-known/agent-card.json + .well-known/mcp.json |
| Machine-readable · llms-full.txt (bulk docs for models) | 149 | llms-full.txt |
| Machine-readable · schema.org product/offer markup | 139 | schema.jsonld |
| Payable · Agent payment protocol referenced | 114 | x402-server.js + AGENTS-API.md |
| Discoverable · llms.txt present | 76 | llms.txt |
| Discoverable · sitemap.xml present | 43 | sitemap.xml |
| Purchasable · Commerce paths open to a plain agent | 37 | robots-snippet.txt |
| Machine-readable · llms.txt useful for commerce | 37 | llms.txt + llms-full.txt |
| Purchasable · Pricing published for machines | 37 | PRICING.md + schema.org Offer |
| Reachability · Homepage fetchable | 17 | — |
| Permissive · AI agents allowed by robots.txt | 5 | robots-snippet.txt |
The pattern is exact: the cheap, documentation-shaped checks are widely passed, and the expensive, money-shaped checks are universally failed. That is what a market looks like before someone builds the boring half.
agent finds them ✅ llms.txt, sitemap, fast pages
agent reads the offer ⚠️ prose docs, rarely schema.org Offer, prices behind a signup
agent tries to buy ❌ "Sign up" / "Contact sales" / "Start free trial"
agent tries to pay ❌ no 402, no agent credential, no wallet, no protocol endpoint| Metric | Value |
|---|---|
| Accounts on the ledger | 2 |
| Credits sold | $21.00 |
| Unused credit (our liability) | $19.93 |
| Agent calls metered | 267 |
| Gross billed to agents | $1.07 |
| Our revenue (take rate 15%) | $0.16 |
| Owed to vendors | $0.91 |
Pilot data from live vendor accounts.
We generate and deploy the missing half: 11 distinct artifacts, produced in one command (node build.js <domain>) and verified by the same auditor that produced this report. A vendor running our surface scores 100/100 — publicly checkable at http://acme-data.91.99.165.95.sslip.io/.
Commercial shape: $15,000 to ship the surface, then 15% of what agents actually spend through it, with the first $10,000 free. We are merchant of record, so we carry the tax, disputes and settlement — the vendor keeps the customer and the relationship.
Cited because they are the reason this is urgent, not because they flatter us:
The gap between the fourth and fifth bullets is the entire opportunity: enormous routed volume, almost no completed machine payments.
machine-readable than average. The true population scores lower, not higher.
/api on the audited origin. A company serving a 402 challenge at adifferent path would be under-scored, not over-scored — the failure direction is deliberate.
to take money. Two sites in an earlier pass were reported "agent-ready" on those signals alone while still gating purchase behind a signup; that logic bug was found and fixed, and the corpus re-audited.
see; we report what an external agent meets.
lib/audit.js (WEIGHTS); every raw signal is in teardowns/*.json for anyone who wants to re-score differently.
re-audited afterwards. See PROOF.md.
Next report: 2026-10. To be measured, submit a domain at http://agentcheck.91.99.165.95.sslip.io — it is free and the score is public either way.
Paste a domain → · or see the twelve generated files for any company: /demo/<domain>