← All editions

Autopilot Pulse · #05 · 2026-09-03

Your website has two readers now

Edition #5 cover

TL;DR

For most of the web's life it had one kind of reader: a human, with eyes, on a page. That assumption is quietly breaking. Your site now has a second reader — an AI agent, arriving with a task, reading your HTML as data, and deciding in seconds whether it can use you or should move on. Here's the shift, what I did about it on this very site, and the playbook to do it yourself.

1 — Two infra giants built the same scoreboard

This year both Vercel and Cloudflare shipped a 0-to-100 agent-readiness score — Vercel's `is-agentic` (powered by Ora / Era Labs), Cloudflare's at isitagentready.com [HIGH]. When the two companies that route a huge share of the internet build the same scoreboard in the same year, it isn't a gimmick — it's a signal about where traffic is going.

They don't agree on what to measure. Ora and Vercel are behavioral — they send a real agent to complete a task on your live site and watch where it gets stuck. Cloudflare is protocol-first — it checks whether you speak the emerging standards (Web Bot Auth, Content Signals, MCP Server Cards, agentic commerce). GEO tools like Glippy and aeojs check AI-search visibility. A 100 on one is not a 100 on the others — which is exactly why a consensus view matters.

2 — The day of fixes (68 → 90)

is-agentic score before and after: 68 'Important blockers remain' to 90 'Strong technical baseline'

I ran it on this site and got 68/100 — "important blockers remain." Most of the gap closed in an afternoon, in public: a sitemap and an llms.txt with a "when to use this" section, JSON-LD so an agent can parse who I am, real About / Contact / Pricing pages (agents check these to decide you're legitimate), 404s that actually return 404, and markdown content negotiation.

That got me into the 80s. The uncomfortable framing from Cloudflare's own data: across the 200,000 most-visited domains, only ~3.9% support markdown negotiation and ~4% declare AI usage preferences [HIGH]. Most of the web is invisible to the second reader — so being early here is cheap and rare.

3 — Read vs use: the part most sites miss

The check I couldn't fake was the one that reframed everything: no API, no OpenAPI spec, no MCP server. It made me ask what an index is actually *for*. A human reads it. An agent should be able to call it.

So I built that: a public, read-only REST API, an OpenAPI 3.1 spec, and — the part that matters — an MCP server, so any agent can query the index natively. Security-hardened: read-only, whitelisted fields, validated inputs. The tell on how early this is: Cloudflare found fewer than 15 sites globally publish an MCP Server Card [MED]. Final score: 90/100 — "strong technical baseline." But the 90 isn't the win. The win is that the index went from something an agent can *read* to something an agent can *use*.

4 — What a builder should do about it

Ship the plumbing now — it's cheap and rare. Sitemap, llms.txt with when-to-use, JSON-LD, real trust pages, honest 404s. An afternoon of work that ~96% of the web hasn't done.

Decide what an agent should be able to *do*, not just read. If your product is data — a catalog, an index, a directory, docs — an API + MCP server turns it from a page an agent scrapes into a tool an agent calls. That's a different, more durable relationship.

Don't optimize the badge — optimize the outcome. A 100 doesn't mean an agent can complete a real task on your site. Fix the checks, then watch an agent actually try to use you. The full playbook — every grader compared, the maturity ladder, the tiered steps, and a version you can point Claude Code or Codex straight at — is at autopilotindex.com/guides/agent-ready-web. If you run a business on autopilot, what's the first thing you'd want an agent to *do* on your site — not just read?