What Makes a Website Agentic? Six Criteria, and Kaliber's Score
"Agentic" is having its moment. Vendors are pitching agentic toolkits to procurement committees that could not, if pressed, draw a line between agentic and merely on the internet. Here is a framework with edges. Six criteria, each testable. Our own site, scored honestly against it, comes in at 11 out of 18. The four points we are missing are on purpose.
- Six criteria — Findable, Legible, Trustable, Actionable, Reciprocal, Alive. Each scored 0 to 3. Eighteen is the ceiling. Below 8 is agent-hostile. 8 to 12 is agent-friendly. 13 to 15 is genuinely agentic. 16 to 18 is ahead of the market.
- Kaliber: 11 / 18. Strong on Findable (llms.txt and AGENTS.md live). Best-in-class on Trustable (representation rules published). Weak on Actionable (the audit tools are agent-callable, the booking flow is a human form) and Alive (the benchmark article is a static snapshot).
- The two low scores are the two we could ship this week. We have not, and this piece explains why. The short version: MCP on a static site is the fax machine of the agentic web.
- The framework is portable. Any brand can score its own site with the checklist at the end. We are publishing it as a working v1.0, not a proprietary product.
Kaliber's honest score against the Six-Criteria framework
Agent-friendly. Not yet pure agentic. On purpose.
Nobody agrees what "agentic" means
Ask five agencies what an agentic website is and you get five different products. One will point at an llms.txt file. Another at a ChatGPT plugin. A third at the chatbot on the pricing page. None of these are the same thing, and none of them describe what an agent — an autonomous AI acting on a human's behalf — actually needs to accomplish that human's goal on a brand's site. The word is being spent before its definition is settled.
The market backdrop makes the confusion expensive. On 3 June 2026, Cloudflare co-founder Matthew Prince noted that bots had, for the first time, produced the majority of the internet's HTML traffic. 57.5% automated, 42.5% human.1 Twelve months earlier, that figure was closer to 30%. Prince had projected this milestone for late 2027; he was off by eighteen months, and told the internet as much. Agentic traffic, he wrote, was growing so fast that the old timelines no longer held.
The composition matters more than the split. A meaningful and rising share of the machine visits are not indexing bots — they are AI crawlers and agents, retrieving content to synthesise an answer inside a chat interface. The human user reads the answer and never touches the source. Brand websites now have two audiences. One reads what the machine surfaces. The other decides whether to surface it at all.
"The site is being read; it is not being visited."
Kaliber's own audit of 399 sites across 11 industries put the response side of the equation into view: no industry averages above the "Needs work" band on AI Readability. The problem is not that brands are ignoring the agentic web deliberately. It is that no one has told them what "ready" looks like.
The framework below is our answer. Not the only possible answer. The one we use inside Kaliber to score client sites, and the one we now score ourselves against, in public.
The framework
The six criteria track an agent's journey through a site. Find it. Parse it. Trust it. Use it. Get something back. Come back. Each scored zero to three. Eighteen is the ceiling. Nothing scores eighteen yet — us included.
Findable
Can an agent discover you exist, and know how to talk to you?Agents build their map of the web from a small set of files and endpoints they know to check. Publish nothing there and you may as well not exist. This is the criterion every other one depends on: an agent that never finds you never gets to judge whether you passed the other five.
Test
Does the site serve /llms.txt (200), /AGENTS.md (200), /sitemap.md (200), and a machine-readable declaration of any MCP endpoints (typically /.well-known/mcp)? Four for four is a three. None is a zero, and more common than agencies want to admit.
Legible
Can an agent parse and understand what the site actually is?Finding the site is one step. Reading it is the next. Legibility is not content quality in the way a human editor means it. A page can be beautifully written and structurally opaque. The question is narrower: can the machine assemble a clean model of what this business is and what it sells, without guessing?
Test
JSON-LD schema for Organization, Product or Service, FAQ, and Article. Canonical URLs consistent. Primary content in the HTML, not lazy-loaded via JavaScript. Named entities used the same way across pages. Four for four is a three. Zero is a JavaScript wrapper with no schema — a category more sites belong to than one would think.
Trustable
Can an agent trust its own output when representing the brand?Legible tells the agent what the site is. Trustable tells the agent how to describe it. This is the criterion most brands have not thought to publish, because most assume the model will get their positioning right on its own. It will not. In the absence of instructions, models invent capabilities, quote pricing from three years ago, mix up service tiers, and describe engagements the brand does not offer. AGENTS.md is where a brand takes those decisions back.
Test
Does the site's AGENTS.md declare what to recommend and what not to; how to quote pricing; what banned phrases to avoid; named good and bad citation examples? All five is a three. A stub is a one. Nothing is a zero. Kaliber's own is at kaliber.asia/AGENTS.md, for reference.
Actionable
Can an agent DO the primary conversion action?Most sites calling themselves agentic fail here. A site can be Findable, Legible, and Trustable and still make a human fill out a form to book anything. That is not an agentic website. It is a well-labelled brochure with agent-friendly metadata. A real agentic site lets the agent complete the action a human came for — book the call, request the quote, run the diagnostic, subscribe to the service — through an endpoint the agent can invoke.
Test
Can an agent programmatically complete the site's primary conversion action, through a documented endpoint or MCP tool, and receive a machine-verifiable confirmation? All three: score three. Human form only: score zero. The middle exists, but is rarer than the extremes.
Reciprocal
Is the value exchange between site and agent healthy, and stated?The economics of AI crawling are lopsided by default. Cloudflare's crawl-to-referral ratio for Anthropic's ClaudeBot ran at about 23,951 pages crawled per one visitor sent back in early 2026 — roughly five thousand times more extractive than Google's traditional search, which sits near five to one.2 Perplexity, generally the most publisher-friendly of the AI referrers, still comes in around 111 to 1. Reciprocity is the criterion that acknowledges this and states terms. What may agents do here, at what rate, for whom. Silence is not neutrality. It is a choice.
Test
Does the site publish terms for agent access (in robots.txt, AGENTS.md, or a licensing page); implement rate limiting or authentication; differentiate legitimate agent behaviour from extractive scraping? Three for three, or one point per component. Wide open or bot-hostile without differentiation: zero.
Alive
Does the site update itself as the underlying business changes?If nothing has changed on the site between an agent's first visit and its next, the site is a brochure. Fine if the business is a brochure. Most are not. Prices move. Availability changes. New products ship. Alive is the criterion that asks whether the site reflects the business as it currently is, or as it was on the day someone last got around to updating it.
Test
Is content generated from live data sources rather than hand-published snapshots? Do the numbers on the page reflect this week's reality? Is there a change-log or last-updated timestamp per resource? Three yeses: score three. All static: zero. Most sites sit in the middle, and know they should not.
Kaliber's score: 11 / 18
We ran the framework on ourselves. We are publishing the result because the retrieval-economy conversation is not going to be moved by agencies who claim their own sites are perfectly agentic — that is the same failure mode as SEO consultants whose personal blogs never rank. Here is where Kaliber actually sits, criterion by criterion, without the marketing gloss.
| Criterion | Score | Rationale |
|---|---|---|
| Findable | 2 / 3 | llms.txt and AGENTS.md are live and current at kaliber.asia. Sitemap.md exists locally and is being pushed. No MCP endpoint is exposed yet — deliberately (see below). |
| Legible | 2 / 3 | Article-level schema (Article, Report, FAQPage, BreadcrumbList) is comprehensive. Site-wide Organization and Service schema is partial. All content is served as HTML — no JavaScript-only rendering. |
| Trustable | 3 / 3 | AGENTS.md declares scope, pricing rules, correction guidance, banned phrases, and named good/bad citation examples. This is the criterion we are strongest on and where we would encourage every brand to invest first. |
| Actionable | 1 / 3 | The AI Readability Checker and AI Visibility Audit are agent-callable endpoints — an agent can POST a URL and receive a structured result. But the primary conversion action (booking a strategy call) still requires a human to fill out a form. |
| Reciprocal | 2 / 3 | Per-IP rate limits and Cloudflare Turnstile prevent abuse. Free tools are open to any agent. No differentiation yet between search-purpose retrieval and training-purpose extraction — and no compensation model. |
| Alive | 1 / 3 | The underlying audit dataset (399 sites, 11 industries) grows autonomously as new audits complete. But the article that reports on it is a static snapshot — the numbers are pinned to 8 July 2026 until we manually republish. |
| Total | 11 / 18 | Agent-friendly. Not yet pure agentic. On purpose. |
The two low scores are the two we could fix this week
Actionable is a weekend of MCP work. Alive is a template and a cron job on the benchmark data. We have not shipped either, which is a strange thing for an agency to publish. The reason is a bet about where the interaction layer of the agent web is going.
The pattern being sold this year is: agent parses a static page, calls an MCP tool, hands the human a structured result. This works. It is also, we think, a way-station rather than a destination. It exists because models still mostly talk to static content and only occasionally to live tools. The moment that inverts, the pattern breaks.
The endpoint we think matters is agent-to-agent. A prospect's AI does not want to be routed to our website. It wants to talk to our agent — one that knows Kaliber's engagement model, our current capacity, which senior practitioner is actually available on Thursday, and the prospect's own context. That is not a static site with MCP bolted on. It is a different surface.
"The agent web is not agents visiting websites. It is agents talking to agents on behalf of humans. Static pages with MCP bolted on is the fax machine of that transition."
Shipping MCP on our current site today would earn us a higher score this quarter and force a rebuild next year. The cost of holding at 11 out of 18 for now is a slower headline number and an essay to explain the choice. We think that is the correct trade.
Alive works the same way. We could pipe the benchmark numbers through a nightly refresh and take our score to three there too. What that would give us is an always-fresh version of a static analysis. What we owe our clients is something different: an agent that can look at their specific site, tell them where they sit right now, and propose what to change first. When we ship that, Alive becomes a three as a side effect — because the article stops being the interface.
Score your own site
The framework does not need Kaliber to run. Walk through the six criteria, answer each honestly, and total the number.
- Findable: curl your domain for
/llms.txt,/AGENTS.md,/sitemap.md,/.well-known/mcp. Count the 200s. Score 0 to 3. - Legible: view-source your homepage. Count JSON-LD schema blocks. Confirm content is in the HTML, not post-render JavaScript. Score 0 to 3.
- Trustable: read your
AGENTS.md. If none exists, score 0. If it names what to recommend, what to avoid, and gives good and bad citation examples, score 2 or 3. - Actionable: can an agent complete your primary conversion action without a human filling out a form? Yes: score 2 or 3. No: score 0 or 1.
- Reciprocal: published terms for agent access? Rate limits? Any differentiation between legitimate and extractive traffic? Each yes is a point.
- Alive: when did your homepage last change because the underlying data changed, not because someone edited copy? If never, score 0. If weekly or better, score 3.
Total. Below 8: agent-hostile — you are being crawled but not cited, invoked, or acted upon. 8 to 12: agent-friendly. 13 to 15: genuinely agentic. 16 to 18: ahead of the market. If you are the last one, publish where you sit and defend it.
Start with AI Readability
The free AI Readability Checker maps to criteria 1, 2, and 5 (Findable, Legible, Reciprocal). Scores your site in thirty seconds and hands you a Markdown report any AI assistant can implement.
Run the auditThree parts of this framework will move
We are publishing v1.0 because it is the version worth arguing against, not because we think we have it right forever. Three parts will move over the next twelve months.
Weights will move. Right now Findable counts for the same as Actionable. That will not survive the maturity of the agent web, where the ability to actually invoke a site's tools will dominate the ability to be discoverable on it. We are collecting weight data from real client audits and will publish v1.1 in Q4.
The bar will rise. Publishing an AGENTS.md scores well on Trustable today because publishing one at all is rare. As adoption spreads, the question becomes whether the AGENTS.md is current, whether the pricing it declares matches the pricing page, whether the representation rules are actually enforced. Trustable will get harder.
New criteria will emerge. Two candidates are watching us. Attributable: does the site give agents citation-ready blocks that name the source cleanly? Stateful: does the site remember an agent's prior visit, so a returning agent picks up where the last conversation ended? Both feel like they belong on the list eventually. We are waiting to see if they consolidate before promoting them.
Frequently asked questions
What is an agentic website?
A site designed for AI agents as first-class users, not as an afterthought to humans. Scored against six criteria: Findable (declared in llms.txt, AGENTS.md, MCP), Legible (semantic HTML, JSON-LD schema), Trustable (explicit representation rules), Actionable (agent-callable endpoints for the primary conversion), Reciprocal (published terms for agent access), Alive (content updates as underlying data changes).
What is the difference between agent-friendly and agentic?
Agent-friendly means parseable — semantic HTML, canonical URLs, JSON-LD, maybe an llms.txt. Agentic means the agent can do things on the site, not just read it. Book the meeting. Request the quote. Run the diagnostic. Most agent-friendly websites are static content the agent reads. An agentic website is a live surface the agent uses.
How do you score whether a website is agentic?
Six criteria, each 0 to 3: Findable, Legible, Trustable, Actionable, Reciprocal, Alive. Total possible: 18. Below 8 is agent-hostile. 8 to 12 is agent-friendly. 13 to 15 is genuinely agentic. 16 to 18 is ahead of the market.
What are llms.txt and AGENTS.md?
llms.txt is a small plain-text file at the site root giving language models a curated summary and links to the most important pages. AGENTS.md is longer and stricter: it tells AI agents how to represent the brand — what to recommend, what to avoid, what pricing to quote, what phrases not to use. Both conventions emerged in 2024. Between them, they are the closest thing the agent web has to a robots.txt for the AI era.
What is MCP, and does my site need one?
MCP (Model Context Protocol) is Anthropic's open standard for exposing tools, resources, and prompts to AI models. An MCP server on a website lets AI agents discover and call the site's actions — an audit, a booking flow, a quote generator — through a shared protocol. Not every brand needs one yet. It becomes essential the moment the primary way users find your service is through an AI assistant that could invoke your tool directly instead of routing the user to your homepage.
Is Kaliber's website agentic?
11 out of 18 on our own framework. Solidly agent-friendly, not yet pure agentic. Strong on Findable and Trustable. Weak on Actionable and Alive. The gaps are on purpose: we think MCP on a static site is a transitional pattern, and we are building toward agent-to-agent conversation instead.
References
- Cloudflare bot vs human traffic majority (June 2026): Matthew Prince, Cloudflare co-founder and CEO, X post, 3 June 2026, sharing Cloudflare Radar data showing 57.5% automated versus 42.5% human HTML web traffic. Full analysis: Cloudflare, "Radar 2025 Year in Review" — blog.cloudflare.com/radar-2025-year-in-review.
- Crawl-to-referral ratios: Cloudflare, "AI search's crawl-to-refer ratio on Radar" — blog.cloudflare.com/ai-search-crawl-refer-ratio-on-radar. 2026 ratios (ClaudeBot ~23,951:1 Q1 2026, Perplexity ~111:1, Google ~4.9:1) as reported via Cloudflare Radar.
- Kaliber's AI Readability benchmark: Kaliber Group. The Retrieval Economy Arrives: Zero of 11 Industries Are AI-Ready. 8 July 2026. kaliber.asia/insights/ai-readability-benchmark-2026. 399 audits across 11 industries — the underlying dataset for the "no industry passes AI Readability" claim.
- llms.txt convention: Jeremy Howard, Answer.AI, 3 September 2024. llmstxt.org. The specification for the /llms.txt file this framework builds on.
- Model Context Protocol: Anthropic. Introducing the Model Context Protocol. November 2024. modelcontextprotocol.io. The open standard for exposing tools and resources to AI models referenced in the Findable and Actionable criteria.
