We Scored 124 Dev Tools on Agent Readiness. It Barely Predicts Who Agents Pick.

Cloudflare shipped an Agent Readiness score during Agents Week: a free scanner that checks whether your site publishes the machine-readable furniture AI agents look for. Markdown for agents. Link headers. An API catalog. An MCP server card. It grades you 0 to 5, and it now feeds Cloudflare Radar and URL Scanner.

It is a good tool measuring a real thing. It also arrives with an implied promise that nobody has tested: that being agent ready has something to do with agents choosing you.

We can test it. We already measure whether agents pick a tool over its competitors on unbranded tasks. So we ran the readiness scanner against all 124 domains on the Pickrate leaderboards and joined the two datasets.

The correlation between agent readiness and getting picked is 0.19. That is close enough to nothing that you should stop treating readiness as a growth lever.

What we did

124 unique vendor domains, every one on a Pickrate leaderboard, scanned through the public Agent Readiness API. 121 of them joined cleanly to a measured Pick Rate across 22 categories. Pick Rate is measured on unbranded prompts, so no vendor name ever appears in the question. Readiness is Cloudflare's level, plus the finer grained count of individual checks passed out of 16 scored (their five commerce checks don't count toward the level, so we excluded them too).

Four domains never got a score, and they are worth naming: akamai.com, make.com, platform.openai.com, and vonage.com all returned a 403 bot challenge to the scanner. A site that blocks a readiness scanner blocks agents. That is a more decisive answer about the agent channel than any score.

Finding 1: almost nobody is agent ready

Before the correlation, the distribution. Out of 124 developer tools, the kind of infrastructure companies whose entire customer base is engineers:

Eleven vendors out of 124 are above Level 2. Sixty eight percent sit at Level 1, which means they have a robots.txt and a sitemap and essentially nothing built for agents.

The per check numbers show exactly where the line falls. Sitemap: 94%. robots.txt: 88%. Explicit AI bot rules: 83%. Then the floor drops out. Markdown for agents: 17%. Link headers: 13%. Agent skills index: 11%. API catalog: 10%. MCP server card: 8%. auth.md: 3%. WebMCP: 2%. DNS-AID and A2A agent cards: zero out of 120.

Everything invented before roughly 2023 is near universal. Everything invented since is near zero. That is not a story about which standards are better designed.

Finding 2: readiness barely moves with Pick Rate

Across 121 vendors, correlating Cloudflare's level against measured Pick Rate gives a Spearman coefficient of 0.187 (Pearson 0.236). Using the finer grained share of checks passed instead of the level gives 0.180 (Pearson 0.151).

A cleaner test, because it controls for how hard each category is: in each category, does the most agent ready vendor also win it? Seventeen categories had any spread in readiness at all. The readiest vendor won 8 of 17. That is a coin flip.

Mean Pick Rate by level does not climb either:

Level 4, the second most agent ready tier, has the same mean Pick Rate as Level 1. If readiness were driving selection, that row could not look like that.

Finding 3: the outliers are the whole argument

Averages hide it. The individual pairs do not.

Stripe scores Level 1, passing 4 of 16 checks. No markdown for agents, no Link headers, no API catalog, no MCP server card. It wins 79.6% of payments trials, the single most dominant number on our leaderboards. The entire payments category sits at Level 1, and the category still has a runaway winner.

Telnyx scores Level 5, passing 14 of 16. It is one of the three most agent ready companies we measured, ahead of nearly every tool in this corpus. Its Pick Rate in SMS is 0.0%. Twilio, at Level 2 with 4 of 16, wins the category.

The same shape repeats:

• Resend (L5, 10/16) loses transactional email to Postmark (L1).
• Supabase Auth (L4) gets 0.3% and loses auth to Auth.js (L1).
• Supabase Storage (L4) gets 0.0% while Amazon S3 (L1, 3/16) takes 35.1%.
• GitHub Actions scores L0 with 2 of 16 and wins CI/CD at 28.6%.
• Netlify and Render both sit at L4 in hosting. Vercel, at L2, beats both.

There is a pattern in who wins anyway, and it is not subtle. Stripe, S3, GitHub Actions, Linear, Prisma, PostHog. These are tools with a decade of documentation, tutorials, blog posts, and Stack Overflow answers sitting in the pretraining corpus. The model already knows them. It does not need to fetch anything to have an opinion.

What we think is actually happening

Agent readiness is a retrieval time signal. It matters in the moment an agent fetches your site, reads your markdown, and calls your MCP server.

Pick Rate today is dominated by a training time prior. When someone asks an agent how to accept payments, the answer is Stripe before any page is fetched. The readiness signal never gets consulted, because retrieval never happens.

That framing explains the outliers cleanly. Telnyx did everything right for the agent that shows up. Not enough agents show up, because the model already had an answer. Readiness shapes the encounter; it does not create it.

Where we could be wrong

The honest limits, since we would rather state them than have them found:

The top of the scale is thin. Only 11 of 124 vendors are above Level 2: two at L3, six at L4, three at L5. Any conclusion about the high end rests on small numbers, and one or two vendors moving would shift those means noticeably.

This is a snapshot, not a time series. Most of these standards are months old. Pick Rate reflects years of accumulated mindshare. Finding no correlation today is not evidence that readiness will not matter in a year, and if agent mediated traffic keeps growing it probably will. What the data rules out is the claim being made right now, that publishing these files moves your selection rate.

Correlation is not causation in either direction. It is equally possible that being unpopular gives a company the free time to implement every emerging standard, and that scrambling for the agent channel is a symptom of losing the human one.

So should you bother?

Yes, but be honest about which problem it solves.

Readiness is necessary but not sufficient, and right now it is not even necessary. What it buys you is that when an agent does arrive, it can read you, call you, and get an accurate answer instead of parsing 200KB of markup or hitting a bot challenge. That is worth an afternoon. It is table stakes, and table stakes are cheap.

What it does not buy you is the encounter itself. If agents are not reaching for you, the fix is upstream of your .well-known directory: being present and correct in the material models learn from, and in the search surfaces they retrieve from.

The uncomfortable version, for anyone selling agent readiness as a growth strategy: on 121 measured tools, the most agent ready vendor won its category slightly less than half the time. Publish the files. Then go do the hard part.

Every readiness score in this post is live on the tool pages, next to the Pick Rate it failed to predict. Browse the leaderboards, or look up your own tool.