Manifesto
The web has a second audience
Machines now read the web at scale: crawlers feeding models, assistants composing answers, agents running errands. We build for that audience, because being legible to it is becoming as decisive as being found ever was.
Dieses Dokument ist in englischer Sprache verfasst.
The readers changed
The busiest new readers of the web are not people. Cloudflare measured OpenAI's crawler growing 305 percent in requests in the year to May 2025, with AI crawlers taking an ever larger share of all verified bot traffic, and the curve has kept climbing since.
These readers do not browse; they ingest. Cloudflare's crawl-to-refer data shows AI platforms reading hundreds to thousands of pages for every human visit they send back. For a growing share of your audience, the crawl is the visit: the answer gets written wherever your content was legible, with or without you.
The numbers, from the source: Cloudflare: who is crawling your site (2025) · Cloudflare Radar: the crawl-to-refer ratio
Sites still dress for eyes
Most sites are built for human vision: meaning carried by layout, content that only exists after JavaScript runs, no declared structure, no discovery files, no stated permissions. A person squints and copes. A machine reads a blurry page, or an empty one, and moves on. Nothing fails loudly; you are simply absent from answers and errands you never see.
What we believe
Four positions, each held with evidence rather than mood:
- Machine readability is measurable. Not vibes: named checks with public weights, scores that mean the same thing on every site they touch.
- It is maintained, not achieved. Sites drift, templates change, standards move. A score is a photograph; only the series tells the truth, so the series is the product.
- Proof beats markup. The test of an agent-ready site is an agent finishing an errand on it, so we send a real one and publish the replay.
- The ruler must be public. Our engine's checks are documented and versioned, and we run them against our own site first; a measurement you cannot audit is an opinion.
What we built
One loop, kept short:
- Measure: 26 checks over every page template of a site, not just the home.
- Watch: history per check, scheduled rescans, drift disclosures, and alerts on rules you write.
- Fix: a remediation guide per check, files generated from your own pages, one-click autofix artifacts, and industry plugins that ask what your kind of site owes an agent.
- Prove: real agent task replays on your site, published at your discretion.
- Open: the same measurements over an HTTP API and an MCP server, so the gate lives in your CI rather than in a browser tab.
Where this goes
The agentic layer is compounding and its conventions are young: llms.txt, MCP server cards, agent cards, payment and identity rails still being drawn. We do not claim to know which of them win. Our commitment is narrower and more durable: keep an honest, versioned ruler over whatever the machine-readable web becomes, and keep the distance between knowing and fixing short.