Cloudflare's Agent-Readiness Checker Claims to Check for llms.txt. It Doesn't.

We ran RCTV and five AI-video vendors — Higgsfield, Runway, Luma, Google, Kling — through a tool that claims to check for llms.txt. It doesn’t.

isitagentready.com is a free scanner Cloudflare built to grade how ready a website is for AI agents. Its own meta description says exactly what it promises: “Check for llms.txt, MCP, agent skills, and other agent-friendly standards.” We scanned six domains against that promise. Not one of the tool’s 20 individual checks looks for an llms.txt or ai.txt file — not RCTV’s own, live and substantial at rctv.com/llms.txt, and not anyone else’s. The scores below are real. What they measure is narrower than the tool’s own pitch, and that gap is the actual finding.

The scores

All six sites land on the scanner’s lowest badge, “Level 1 — Basic Web Presence.” RCTV and Google — measured at the deepmind.google domain level, not a Veo-specific page — tie for last at 21/100. Higgsfield, Runway, Luma, and Kling all land at 29/100.

Twenty checks, five companies with real engineering teams and stated AI-agent ambitions, and the rubric can’t tell four of them apart. Whatever actually separates a Higgsfield from a Kling on agent readiness, this scanner isn’t the instrument finding it.

Worth stating plainly: isitagentready.com is a Cloudflare product, scoring the web against standards Cloudflare itself put forward — Web Bot Auth, which Cloudflare announced in May 2025, and the Content Signals policy it introduced that September. That doesn’t disqualify the tool. It does mean the standards it rewards aren’t neutral consensus — they’re the ones one vendor is currently pushing hardest.

What the checker actually checks

The scanner’s 20 checks span Discoverability, Content, Bot Access Control, API/Auth/MCP/Skills, and an optional Commerce category nobody here triggers. None of them checks whether an llms.txt or ai.txt file exists at all. The only path either file has to earn any credit at all is indirect: an HTTP Link response header on the homepage, carrying a relation type the tool recognizes as agent-useful.

Higgsfield actually tries this. Its homepage serves Link: </llms.txt>; rel="llms-txt", </llms-full.txt>; rel="llms-txt"; title="Full" — a real, working pointer aimed straight at its own llms.txt files. The scanner parses it, finds four link relations, and its own trace concludes: Found 4 link(s) but none with agent-useful rel types. Zero credit. The one site in this set that engineered a machine-readable signal to its own llms.txt scores identically on Discoverability to the ones that never tried.

Runway, meanwhile, picks up a Discoverability point its five peers don’t — entirely from a Link header carrying rel="alternate"; hreflang="ja", ordinary multilingual-SEO routing that predates “agent-ready” as a phrase by a decade. The scanner’s own log flags the relation type as agent-useful — the same language it withheld from Higgsfield’s llms.txt headers above. A purpose-built agent-discovery file, explicitly offered, doesn’t qualify; a language-alternate tag does.

RCTV’s own llms.txt and ai.txt are both real: citation preferences, an AI-native editorial-policy statement, a machine-readable index of every load-bearing page on the site. RCTV’s homepage doesn’t serve the Link header Higgsfield does, so the files were never going to register through that path either. The honest read isn’t that RCTV got cheated out of points — it’s that the checker has no mechanism, working or broken, that would have counted this specific investment either way.

The one real difference

Not everything here is a rubric quirk. Kling is the only one of the six sites with any credit at all in the API/Auth/MCP category — one point out of seven, but a genuine one. Kling’s homepage serves a working /.well-known/oauth-protected-resource endpoint; the scanner’s trace shows a clean 200 response reading Valid metadata with resource: https://kling.ai/mcp. That’s real infrastructure — a discoverable, standards-compliant entry point for an AI agent to authenticate against Kling’s MCP server, consistent with the MCP and CLI tooling Kling shipped in mid-July — not a checkbox it ticked by accident.

Sit with both findings together. The same scanner, on the same day, caught something real at Kling and missed something real at Higgsfield and RCTV. A tool that’s sometimes right isn’t worthless. It just isn’t the single number “agent-ready” implies it is.

The take

A scanner built by an interested party is still useful. Cloudflare didn’t invent the idea that AI agents need machine-readable entry points, and robots.txt rules, sitemaps, and OAuth discovery are worth having regardless of who’s grading you on them. But “agent-ready” isn’t a single number any more than “SEO-optimized” ever was — a tool built around 20 specific protocol checks will always undercount whatever it didn’t think to check for six months ago.

RCTV’s own 21/100 is the clean illustration. The investment is real: an ai.txt that explicitly welcomes AI training and states a citation preference, an llms.txt summarizing every load-bearing page on the site, a robots.txt that names eleven AI crawlers by user-agent and tells each one what it can do. None of that shows up in this score, and none of it is wrong because of that. It’s a data point about what one popular Cloudflare tool currently measures, not a verdict on whether the underlying work matters. Written down, the distinction sounds obvious. It’s exactly the one a bare leaderboard number erases.

What we’re watching

Two things. Whether isitagentready.com adds a direct llms.txt/ai.txt check — Cloudflare ships new checks to this tool faster than most standards bodies ratify the protocols behind them, and closing the gap this piece describes would be a small fix. And whether Content-Signal directives, the one category nobody in this set has touched, RCTV included, become worth adopting on their own merits rather than for a score that currently gives them no visible return. We’ll re-run this scan the next time either changes.