how it works

A real AI agent visits your site with no context, no rehearsal, and no second chance — the way every agent arrives. It attempts four tasks a buyer's agent would: identify what you sell, find your pricing, locate your docs or help, and reach the point of signup. A judge grades what it understood.

Paste your URL on the homepage and press Run cold read. The read takes a minute or two; you see each check finish, and there's a game to play while you wait. Your email gets you the whole report: score, mode gap, and every fix.

0–100, weighted across the four tasks plus deterministic trust-and-structure checks (llms.txt, structured data, robots policy, content negotiation, meta tags, fetch-mode content, sitemap). 90+ is an A. Every deduction carries a receipt — a transcript excerpt or a check output. No unexplained point losses.

Sites are read two ways: fetch mode (raw HTTP, no JavaScript — how API-based agents read) and browser mode (full render). Content that exists only after JavaScript runs is invisible to fetch agents. The MODE_GAP flag names exactly what they lose. It never double-counts against the score.

Passive observation of what your public surface exposes to a hostile agent — unprotected signup forms, prompt-injection surfaces, extraction overexposure. Coldread never submits, probes, or exploits. Risks lead the fix list and never touch the score.

The report ends with one ranked list of fixes: risks first, then what JavaScript hides, then what's missing, by points returned per unit of effort. Each comes with a starting point and a prompt for your AI coding tool; Copy all takes the lot. The usual first hour: publish llms.txt, name AI agents in robots.txt, add JSON-LD, serve markdown to agents that ask, and make sure your pricing survives JavaScript being off.

Yes. Paste a URL for the grade; your email gets you the whole report.

A pinned Claude model runs the agent and the judging, at temperature 0 for judges, best-of-two on every task — we grade whether your site CAN be understood. The full methodology is published openly; it's the moat, not a leak.