Agent-Adoption Check
Scan any domain to see how it presents itself to AI agents — robots.txt rules, sitemap, llms.txt, AI-bot directives, MCP server cards, A2A agent cards, OAuth discovery, and 18 more checks. Every scan returns a level (Basic Web Presence, AI-Aware, or Agent-Optimized) and a 0–100 score, with a per-check breakdown explaining each result in plain language.
Open specification · CC-BY 4.0 · 25 checks · 4 categories
25 checks across 4 categories
Each category groups a set of HTTP-level signals an agent looks at when it visits your site. Twelve checks count toward the score; thirteen are surfaced for context (the underlying spec is too new, or the check measures presence rather than correctness).
Discoverability
Can agents find your site's signals — robots.txt, sitemap, HTTP Link headers.
Access Control
Are you telling AI systems what they may and may not do — per-bot rules and content-usage directives.
Content Readability
Once an agent fetches a page, can it actually read it — server-side rendering, markdown twins, page sizing.
Agent Endpoints
Have you advertised agent-native interfaces — MCP server cards, OAuth discovery, A2A agent cards.
Where your domain sits, in plain language
The level is the headline; the score is supporting detail. Each level requires the previous one, plus passing one specific gate-check. There's no L4+ in v1 — the agent web isn't there yet.
Basic Web Presence
Your site has the basics agents look for: a robots.txt with at least one User-agent rule, a working sitemap, and honest HTTP error codes. Most of the web sits here.
AI-Aware
You've signaled that you've considered AI: per-bot rules in robots.txt or content-usage directives declaring what AI systems may do with your pages. The L1→L2 gate is the content-signals check.
Agent-Optimized
Agents fetch your pages cleanly: server-rendered HTML or markdown twins, page sizes that fit context windows, and content negotiation that returns markdown when an agent asks. The L2→L3 gate is markdown-negotiation.
The honest version, in marketer language
Before you trust this score, here's what we found when we ran a 908-brand correlation study across Claude, ChatGPT, and Gemini in April 2026.
Agent-readiness is real, but the effect is small.
We studied 908 brands across Claude, ChatGPT, and Gemini. Agent-readiness signals do show up in how often a brand gets cited — but the strongest effect we measured was medium-sized, and most are smaller. Adopting these practices is worth doing on its own merits; expecting them to transform your AI visibility on their own isn't.
Claude and ChatGPT disagree on what makes a good citation.
On four signals — sitemap, OAuth discovery, robots.txt, and markdown negotiation — Claude rewards them and ChatGPT punishes them, or the reverse. The two systems pick brands by structurally different criteria. There is no universal score that wins both.
The agent web is still being built.
Twenty of the sixty-six checks we tested have under five percent real-world adoption: MCP server cards, A2A agent cards, AGENTS.md, OAuth-protected resources. The specs exist; the practice is arriving slowly. We track it as it does.
We published the inconvenient finding too.
Our v1 score didn't predict LLM visibility outcomes. Rather than retire it quietly, we published the result and we're calibrating v2 around the two signals that actually graduated the pre-registered statistical bar.
What a finished scan tells you
When a scan completes you land on a permalink for that domain. The page walks you through, in order:
- Level — the headline. L1, L2, or L3, named (Basic Web Presence, AI-Aware, or Agent-Optimized). What we'd publish about your site.
- Score — a 0–100 weighted pass-rate over scored checks. Informational checks don't count.
- Per-category subscores — discoverability, access control, content readability, agent endpoints. Where your site is strong; where it lags.
- Per-check breakdown — every check with pass/fail/neutral, what it measures, and why it matters.
- Category benchmark — if your domain belongs to a tracked Respectarium category, your rank against peers and the top brands by score.
This scanner measures adoption of agent-readiness practices defined in the Agent-Adoption Specification V1.0. Implementing these practices makes your site more accessible to AI agents. Our research shows agent-readiness signals are a small but real factor in LLM visibility — implementing these checks doesn’t guarantee improved AI ranking, which depends on many factors (brand recognition, training data, content quality) outside this scanner’s scope.
If you reference this in research or journalism
The methodology and the scanner output are open under CC-BY 4.0. Anyone can build their own implementation against the same spec. Use the citation below if you write about a scan or the tool.
Frequently asked
What does this tool measure?
Twenty-five HTTP-level checks across four categories — discoverability, access control, content readability, and agent endpoints. The full check list and methodology lives in the open Agent-Adoption Specification V1.0.
How is the score calculated?
It's a weighted pass-rate over the twelve scored checks. Each scored check carries a weight from 4 to 10 based on impact. We sum the weights of passes, divide by the sum of weights of pass-or-fail outcomes, and multiply by 100. The thirteen informational checks don't count toward the score. Some clusters of checks can also cap the score — for instance, if your homepage is a JavaScript-only single-page-app shell that returns nothing readable to an agent, the final score is capped.
Does a high score mean my brand will be cited by ChatGPT or Claude?
No. Our research found agent-readiness is a small but real factor — medium-sized at best for the strongest signal. AI visibility depends on brand recognition, training data, content quality, and many factors outside this scanner. A high score makes your site more accessible to agents, which is worth doing on its own merits.
Why are some checks marked 'informational'?
Two reasons. Some specs are too new and adoption is too low to score reliably (MCP server cards, A2A agent cards, AGENTS.md, OAuth-protected resources). Some checks measure presence of a signal where the right answer depends on policy rather than correctness (Web Bot Auth directories, content-usage directives). All informational checks are surfaced so you can see them, but they don't move the score.
Can I cite this scan in research or journalism?
Yes. The methodology, the scanner output, and the underlying research are open under CC-BY 4.0 — quote, share, build on them with attribution. A copyable citation block sits above this FAQ; the spec, the correlation study, and per-scan permalinks are stable references.
Where can I read the full spec and the research?
The Agent-Adoption Specification V1.0 is at /spec/agent-adoption/v1. The empirical correlation study (the research that informs how we frame this tool) is at /research/correlation-2026-04, with a plain-language interpretation at /research/correlation-2026-04/interpretation.
Does my domain get stored after I scan?
Yes. Every scan creates a permanent public permalink at /agent-adoption-check/{your-domain}. Re-running a scan reuses that permalink, so links you've shared keep working. The result page is public and crawlable — this tool is part of the open observatory.
How does this compare to Cloudflare's or Fern's scanners?
Different implementations of overlapping ideas. Even where we share check names with another scanner, they can fire on different brand subsets — we measured roughly zero shared variance with one peer scanner on a same-named check. All implementations are valid; we publish ours alongside the research that backs it.