Is your website ready for AI agents?

The web is getting a second audience

Humans browse websites. AI agents are doing the research, the comparison shopping, and increasingly the buying. 36% of people already use AI instead of Google for some searches. 93% of AI-powered searches end without a click to any website. The $10.9B AI agent market is growing at 46% annually.

This isn't coming. It's here. MCP — the Model Context Protocol — went from zero to 10,000+ indexed servers in 18 months. Anthropic donated it to the Linux Foundation. AWS, Google, Microsoft, and Salesforce are backing it. 40% of enterprise apps will have AI agents by end of 2026.

What the score measures

We scan six categories of AI agent readiness. No AI model required — it's all deterministic HTTP checks with soft-404 detection. Takes 2–5 seconds per domain. As of April 2026 we run Algorithm E — “Spread” (v3), which replaces our earlier Binary Boost model. Two big changes matter if you've seen your score before:

1. Web quality now counts more. The web-quality ceiling moved from 20 to 30 points, so a well-built site with semantic HTML, JSON-LD, and fast SSR can earn real credit before any AI signals show up.
2. We probe developer subdomains. Most companies don't put openapi.json or llms.txt at the root of their marketing domain — they put them under developer., docs., developers., or api.. Algorithm E probes those subdomains in parallel and counts a hit anywhere under your canonical brand, plus a small +5 developer-first bonus for the intentionality of surfacing those endpoints at all.

The net effect is a clearer distribution between 0 and 100, with two meaningful ceilings below the top.

The two ceilings

The two horizontal lines in the table aren't arbitrary — they map to two very different states of the modern web.

The 50 ceiling is where a diligent, web-native SaaS company lands today. They've shipped llms.txt (the 2024 year in review for most tech brands), they have JSON-LD for SEO, their marketing site renders server-side, their sitemap is fresh. They look great to humans. They're legible to LLMs. But they haven't done the next thing — they haven't exposed a machine-actionable interface. An agent can read them, but can't transact with them.

The 100 ceiling is true agent-native. It requires at least one of: an OpenAPI spec agents can consume programmatically, an ai-plugin manifest, or — the highest-value signal — an mcp.json at the root or on a developer subdomain declaring what tools an agent can call. Very few companies are here yet. That's the opportunity.

Why I built this

I've spent 15 years building data products — entity resolution, identity graphs, audience platforms. Hundreds of millions of company records, thousands of attributes per record. But none of them had this one.

When I got a Mac Studio M4 Max, Claude Code on Max plan, and access to open-source models through Ollama, something clicked. I could build real data products — the kind that used to require a team and a budget — by myself, overnight, at zero inference cost. My co-founder Jody and I started Product Hacker as a place to do exactly that.

Honestly, this is the most exciting thing that's ever happened to me professionally. I wake up every day energized. I crush my day job. And then I can't wait to get home and see what Stu — our Mac Studio — built overnight. 8,250+ businesses classified, enriched, and scored with Algorithm E while I slept. It still feels like magic.

The agent readiness scanner was one of those overnight projects. Define the attribute. Write the scanner. Run it against 2,681 domains. Publish the results. All in one night. The v1 algorithm became v2 (Binary Boost), and in April 2026 we shipped v3 (Algorithm E — Spread) with developer-subdomain probing and a web-quality-weighted ceiling. 8,250+ domains scored and counting. No model needed — just HTTP checks and a clear thesis about what matters.

The soft-404 problem

Some sites return HTTP 200 for every URL — including paths that don't exist. An agent hits /mcp.json and gets a 200 with an HTML page instead of JSON. That's worse than a 404 — it's actively misleading. We detect this with a canary URL check and penalize accordingly. If an agent can't tell your real endpoints from fake ones, you're not agent-ready.

What you can do with it

If you're a company: check your score. If you're below 20, you have zero AI integration — agents can't interact with you beyond reading your homepage. Adding an llms.txt file to your root alone is +15. JSON-LD schema in your HTML is +5. An openapi.json on developer.<domain> is another +15 + a +5 developer-first bonus. Thirty minutes of work takes a typical site from 22 (below the 97th percentile) into the 50-60 range (top 3%). An mcp.json manifest adds another +25 and puts you in the top 0.3% with Digits, Checkly, Didomi.

If you sell to companies: this is a segmentation you can't get anywhere else. Companies scoring 40+ have real AI integrations and are your early adopters. Those at 15–25 have solid websites but no AI endpoints — they need education. Below 15 aren't ready for what you're selling yet.

If you're building AI agents: use the score to find businesses your agent can actually work with. No point sending an agent to a site with no structured data, no API, and a JavaScript-rendered splash page.

It's free and open

The Agent Ready 100 ranking is public. The business database is queryable through our MCP server. The scoring methodology is right here on this page. We're building a community data co-op — not a walled garden.

If you want your score, want help improving it, or want to talk about agent readiness for your business, reach out: [email protected]