What is an agent-ready website?
An agent-ready website is one that AI agents can reliably read, act on, and trust. When an autonomous agent — a shopping assistant, a research tool, a browser agent completing a task for its user — lands on an agent-ready site, it can parse what the site says (machine-readable structure), do what it came to do (explicit, stable action pathways), and verify that what it found is current and genuine (evidence and temporal signals).
Agent-ready design sits alongside human UX and SEO rather than replacing them. The same site keeps serving people; it additionally exposes its content and capabilities in forms that software can consume without guessing — semantic HTML, Schema.org JSON-LD, llms.txt, markdown mirrors, and declared capabilities such as agent-protocol manifests.
Where does the term come from?
“Agent-ready” converged from three directions in 2026. On the practitioner side, Chrome for Developers shipped an agent-ready toolkit and Cloudflare introduced an Agent Readiness score. On the academic side, the first controlled study — “Designing Agent-Ready Websites for AI Web Agents” (Elnaffar & Rashidi, July 2026) — formalised the agent-ready website as a design framework and measured what it is worth.
Agent Ready approaches the same property from the measurement side: agent readability is the degree to which a site can be discovered, parsed, understood, and acted on by agents, expressed as a 0–100 score. An agent-ready website is the state; agent readability is how you measure your distance from it.
What does the research show?
The Elnaffar & Rashidi study built two versions of the same prototype shop — a conventional, human-oriented baseline and an agent-ready variant with JSON-LD, exposed structured data, explicit action labels, and temporal-validity fields — and ran three browser agents (GPT-4.1, Gemini 2.5 Flash, Grok-4 Fast) through five tasks, 300 runs in total. The differences were large:
- Strict task success rose from 49.3% to 89.3% — 134 of 150 runs passed on the agent-ready variant versus 74 of 150 on the baseline, with partial completions collapsing from 43 to 3.
- Agents needed 30.4% fewer steps on average (9.31 down to 6.49) to finish the same tasks.
- Token consumption fell between 18% and 40% depending on the model — agent-ready pages are literally cheaper for agents to use.
- The biggest gains came on comparison and multi-constraint selection tasks, where success improved by roughly 77 percentage points each.
Honest caveats: this is a preprint, the site was a prototype built by the authors, two of the five tasks showed no statistically significant improvement, and the study did not isolate which individual intervention did the work. Treat it as a controlled proof of concept — the first quantified evidence for what practitioners had argued from experience — rather than a final word.
What are the three dimensions of agent readiness?
The study organises agent-ready design into three dimensions. They map directly onto artifacts you can publish today:
| Dimension | The question it answers | What the study changed | What you publish |
|---|---|---|---|
| Agent interpretability | Can an agent parse and understand the page? | Product data exposed as JSON and JSON-LD instead of being buried in JavaScript | Semantic HTML, Schema.org JSON-LD, llms.txt, markdown mirrors |
| Agent executability | Can an agent find and perform the right action? | Explicit identifiers, aria-labels, and data attributes on every interactive element | Stable, labelled controls in-page; declared capabilities via MCP server cards, A2A agent cards, or agents.json |
| Agent decision reliability | Can an agent trust what it just read? | Evidence pages and temporal-validity fields alongside the product data | Visible last-updated dates, canonical URLs, consistent metadata, an agent-permissions policy |
One scope note in fairness to the study: its experiment covered the first two dimensions inside the page itself. Site-wide artifacts such as llms.txt and agent-protocol manifests extend the same principles — interpretability and executability declared at the site level rather than per page — but were not part of the experiment, so the measured numbers do not speak to them directly.
How is agent-ready different from SEO and AEO?
Three layers, each building on the last:
- SEO optimises for being found: crawlers index your pages, a ranking decides visibility, and a human clicks the link.
- AEO (answer-engine optimisation) optimises for being cited: an AI assistant synthesises an answer and names its sources.
- Agent readiness optimises for being used: an autonomous agent completes a task on your site — extracts the fact, compares the options, fills the form, calls the API.
The layers overlap in practice. Structured data does triple duty — rankings, citations, and task success — and Google’s own guidance on building agent-friendly sites reads like an SEO checklist extended one layer down, to the point where an agent acts instead of a person.
How do I make my website agent-ready?
- Expose real content in crawlable HTML with JSON-LD. If the answer only exists after JavaScript runs, many agents never see it. Semantic markup plus Schema.org JSON-LD is the single highest-leverage change — it is also what carried most of the measured improvement in the study.
- Publish the discovery files. An llms.txt tells agents what matters on your site; markdown mirrors of key pages give them a clean-text version to read.
- Check your robots policy admits the agents you want. Many sites block AI crawlers by accident — or block the citation crawlers while trying to block trainers. The AI robots.txt generator makes the distinction explicit.
- Declare your capabilities. If your site offers tools, APIs, or agent-to-agent endpoints, say so in machine-readable form: an MCP server card, an A2A agent card, or an agents.json projected from your OpenAPI spec.
- Scan, fix, and re-scan. A free scan grades every layer — site discovery files, per-page extraction signals, llms.txt conformance, and protocol manifests — and each failing check comes with a plain-English fix.
How agent-ready is the web today?
Not very. Across the 1,183 distinct sites scanned on agent-ready.dev as of August 2026, the mean agent-readability score is 56 out of 100, and only 8% of sites reach the excellent band (90–100). Agent-protocol manifests — MCP server cards, A2A agent cards, agents.json — remain in low single-digit adoption. And that sample skews optimistic: it is sites whose owners cared enough to run a check.
The full aggregate — adoption rates, the most common failures, and how often sites block AI crawlers — is published as The State of Agent Readability, updated continuously from the scan corpus. The gap between the agentic-web conversation and what sites actually publish is the clearest signal in the data — and the opportunity, if the measured task-success numbers above hold up in the wild.
Frequently asked questions
- What is an agent-ready website?
- An agent-ready website is one that AI agents can reliably read, act on, and trust. In practice that means three things: machine-readable structure (semantic HTML, Schema.org JSON-LD, markdown mirrors) so an agent can parse what the site says; explicit, stable action pathways (labelled controls, declared capabilities such as MCP server cards or an agents.json manifest) so an agent can do things rather than just read; and verifiable, current information (visible dates, canonical URLs, consistent metadata) so an agent can trust what it found. Agent-ready design sits alongside human UX and SEO — it does not replace either.
- Is an agent-ready website the same as agent readability?
- They are two views of the same property. Agent readability is the measurable degree to which a site can be discovered, parsed, understood, and acted on by AI agents — Agent Ready expresses it as a 0–100 score. An agent-ready website is the state you are aiming for: a site whose agent readability is high enough that agents complete tasks on it reliably. Scan a site and you measure its agent readability; fix what the scan finds and you get an agent-ready website.
- What evidence is there that agent-ready design works?
- The first controlled experiment was published in July 2026 (Elnaffar & Rashidi, arXiv:2607.12056). The authors built two versions of the same prototype shop — a conventional human-oriented baseline and an agent-ready variant with JSON-LD, exposed structured data, explicit action labels, and temporal-validity fields — and ran three browser agents (GPT-4.1, Gemini 2.5 Flash, Grok-4 Fast) through five tasks, 300 runs in total. Strict task success rose from 49.3% to 89.3%, average steps fell 30.4%, and token consumption fell 18–40% depending on the model. It is a preprint based on one prototype site, so treat it as a proof of concept rather than a final word — but it is the first quantified evidence for what practitioners had argued from experience.
- How is an agent-ready website different from an SEO-optimised website?
- SEO optimises for being found: crawlers rank your pages and a human clicks a link. Answer-engine optimisation (AEO) optimises for being cited: an AI assistant synthesises an answer and names its sources. Agent readiness optimises for being used: an autonomous agent completes a task on your site — extracts the right fact, compares options, fills the form, calls the API. Each layer builds on the previous one, and some artifacts do double duty: structured data helps rankings, citations, and agent task success at once.
- How do I make my website agent-ready?
- Five steps, in order of leverage. First, expose your real content in crawlable HTML with Schema.org JSON-LD, rather than only behind JavaScript. Second, publish the agent discovery files: llms.txt plus markdown mirrors of key pages. Third, make sure your robots policy actually admits the agents you want — many sites block AI crawlers by accident. Fourth, declare your capabilities in machine-readable form: an MCP server card, an A2A agent card, or an agents.json manifest, depending on what your site offers. Fifth, verify with a scan and fix what fails: each Agent Ready check maps to one concrete artifact or property, with a plain-English fix.
- How many websites are agent-ready today?
- Very few. Across 1,183 distinct sites scanned by Agent Ready as of August 2026, the mean agent-readability score is 56/100 — and that sample skews optimistic, because it is sites whose owners cared enough to check. Only 8% of scanned sites score in the excellent band (90–100), and agent-protocol adoption (MCP, A2A, agents.json and similar manifests) remains in the low single-digit percentages in the wild. The gap between the agentic-web conversation and what is actually published is still wide — which is also the opportunity for sites that move early.