---
title: "What Is Agent Readiness?"
description: "Agent readiness is whether an AI agent can find, read, use, and pay for a product on the public web. A scan checks that. It does not prove a task finished."
canonical: "https://www.agentrail.co/blog/what-is-agent-readiness"
author: "Sidhant Goyal"
published: "2026-10-17T00:00:00Z"
last_updated: "2026-10-23T05:59:19.998Z"
---

# What Is Agent Readiness?

Agent readiness is whether an AI agent can find, read, use, and pay for a product on the public web. A scan checks that. It does not prove a task finished.

![Gauge with a red needle on the lowest mark, on a soft blur](https://images.ctfassets.net/1htfyfnn1o5y/Eok34tdkcb3WOJ9OKbIbt/0a0c07e26af6e4835e22d97bb48cfbd8/hero_image-2.png?w=1600&fm=webp&q=72)

An agent can quote your homepage and still fail to install the SDK, call the API, or finish checkout. Agent readiness is the property you check before you debate models. It asks whether the public product can be found, read, used, and paid for by an agent, on the surfaces you already ship.

This essay defines that property. The [agent readiness leaderboard](https://www.agentrail.co/agent-readiness) runs one public scan against it. The score, and the steps to run a check, live on the [agent readiness checker](https://www.agentrail.co/agent-readiness/checker). What agents do after they arrive is a separate measurement, covered in [how to measure agent experience](https://www.agentrail.co/blog/how-to-measure-agent-experience).

> **Key Takeaways**
>
> Agent readiness is whether an autonomous agent can find a product, read it, use a supported action, and pay when the product sells that way.
>
> It is a property of the website, docs, and app. The same models do better when the pages are built for them.
>
> A public scan records declared surfaces. It does not record whether a real task finished.
>
> Search ranking and citation tracking ask whether a person, or an answer engine, mentions you.
>
> In Cloudflare’s April 2026 crawl, robots.txt was common and agent-specific standards were rare.

## Agent readiness is whether an agent can find, read, use, and pay

Agent readiness is the extent to which an AI agent can do work on a product without a person translating the interface. Four questions cover that work. Can the agent find the product? Can it read the pages, docs, and specs? Can it take a supported action, such as an API call, an SDK install, or an MCP tool? Can it pay, when the product charges?

Those four questions are the ones behind every company on the [agent readiness leaderboard](https://www.agentrail.co/agent-readiness). The scan uses what an agent can reach in public. It does not log in. It does not call a private API. We scan the website, the docs, and the app the way agents see them, with JavaScript and without it.

An outside framework uses a close definition. Elnaffar and Rashidi, in a 13 July 2026 preprint, describe an agent-ready website as one an agent can interpret, act on, and trust when it decides. Their dimensions are interpretability, executability, and decision reliability ([arXiv:2607.12056](https://arxiv.org/abs/2607.12056)). Our four questions add an explicit payments check and keep the unit as the whole product, not one shop page. The shared idea is use.

## The site changes the outcome when the model stays fixed

Readiness is easy to file under “the model will get better.” The controlled comparison says the pages are part of the result.

The same authors built two versions of one prototype shop. Catalog, prices, stock, and shopping flow matched. One version was a human-oriented baseline. The other added structured data, explicit action labels, and fields that say when a fact is valid. They ran three browser agents, GPT-4.1, Gemini 2.5 Flash, and Grok-4 Fast, through five tasks. That is 300 runs. Strict success moved from 74 of 150 runs on the baseline to 134 of 150 on the agent-ready version. Those rates are 49.3% and 89.3%. Partial outcomes fell from 43 to 3. Average steps fell from 9.31 to 6.49. The largest gains were on product-detail extraction, comparison, and selection under several constraints.

Hold the caveat in the same breath. It's a preprint. The site is a prototype the authors built. It isn't a forecast for your domain. Two facts still travel. The models didn't change between arms. The pages did. If the price exists only inside a screenshot, or the docs appear only after a click, you're running the baseline.

## Ranking and citation ask a different question

Search optimization asks whether a crawler can index a page and a person will click. Generative-engine work asks whether an answer engine mentions you. Microsoft’s Agent Experience essay, published 6 October 2026, separates mention from use. Use asks whether the agent then calls the current API, rather than one you deprecated ([What is Agent Experience (AX)?](https://developer.microsoft.com/blog/what-is-agent-experience-ax/)). Mathias Biilmann described agent experience in January 2025 as the experience an AI agent has as the user of a product.

[Agent experience](https://www.agentrail.co/blog/agent-experience) is that use, for agents acting on your product. Contact-center products use the same two letters for the working life of human support staff. That subject has its own literature. We measure the other one.

A readiness scan is also distinct from an eval of an agent you ship. An eval tests an agent you configure, on a task you script. A readiness scan inspects the public product that agent would have to read and call. Teams use both. We sell simulations as their own product, because a scripted task and a public surface check fail for different reasons.

## A scan records what you have declared

Declared surfaces are the cheap check. You can look for `robots.txt`, a sitemap, `llms.txt`, an OpenAPI file, an MCP manifest, or a payments signal without watching an agent work. The inspection answers a real question. Has the product put the needed things on the record?

It cannot answer the next one. A file can exist and still be ignored. A tool can be listed and still return an error the agent cannot act on. A measurement essay updated 8 October 2026 states the limit directly. A declared-surface score is a precondition. It does not predict that an agent will succeed ([How to measure Agent Experience](https://agentexperience.tech/insights/measure-agent-experience/)). Cloudflare’s public score and a rubric you run by hand are different vantage points. Their numbers don't transfer, and neither transfers to a task-completion rate.

We keep the jobs apart on purpose. The leaderboard answers the precondition. [Agent experience metrics](https://www.agentrail.co/blog/how-to-measure-agent-experience) answer what happened on a task. [Agent experience optimization](https://www.agentrail.co/blog/agent-experience-optimization) is how you change the step that failed.

## Most public sites still speak to crawlers

Cloudflare Radar took the 200,000 most visited domains and dropped categories where this question is a poor fit, including redirects and ad servers. The remainder is businesses, publishers, and platforms an agent might actually use. In that set, published 17 April 2026, `robots.txt` was present on 78% of sites. Content Signals, the lines that say whether content may be used for training, for inference, or for search, were present on 4%. Markdown content negotiation, a response to `Accept: text/markdown`, passed on 3.9%. MCP Server Cards and API Catalogs (RFC 9727), counted together, showed up on fewer than 15 sites ([Introducing the Agent Readiness score](https://blog.cloudflare.com/agent-readiness/)).

| Check | Result in the filtered set | Source date |
| --- | --- | --- |
| `robots.txt` present | 78% of sites | 17 April 2026 |
| Content Signals in `robots.txt` | 4% of sites | 17 April 2026 |
| Markdown content negotiation | 3.9% of sites | 17 April 2026 |
| MCP Server Cards and API Catalogs, combined | Fewer than 15 sites | 17 April 2026 |

A second sample is smaller, and it leans toward people who asked to be scanned. Across 1,183 distinct sites on agent-ready.dev as of August 2026, the mean agent-readability score was 56 out of 100. Eight percent scored from 90 to 100. The publisher says the sample skews high ([What is an agent-ready website?](https://agent-ready.dev/what-is-an-agent-ready-website)). That figure isn't an Agentrail score. It's a second instrument, on a self-selected set, pointing at the same gap. A live homepage and a product an agent can work through are still far apart.

Cloudflare also measured a format effect on its own docs. Serving markdown instead of HTML cut tokens by up to 80% in some cases. That number is theirs, on their pages. It's why “can an agent read you?” belongs in readiness, next to color and layout.

## The scan is limited to the public contract

The leaderboard states the rule. Every company gets the same public scan. If an agent on the public web cannot see a control, that control is outside the score. Grades on the board are words for that public result. A is strong. B is good. C is fair. D or F means the public site is hard for an agent to find, read, or use. A+ sits at the top of that scale. The page does not publish numeric cutoffs for those letters. We will not invent them.

That's why a logged-in app can look finished in your staging review and still score poorly. The agent that arrives from a docs link, an SDK page, or an MCP card doesn't open your admin session. Readiness, in this sense, is the public contract.

Weekly scans keep the contract from going stale. You can also scan now. A finding comes back with evidence and a fix prompt, so a miss is a concrete artifact. The point weights, the letter next to your number, and the click-path for a check are documented on the [checker](https://www.agentrail.co/agent-readiness/checker). This page stops at the definition.

## Readiness sits in front of production traffic

Agents already cross surfaces inside one task. A job can start in the docs, inspect an SDK, call an MCP tool, hit an API, and stop on authentication. Conventional product analytics often stores those as separate events. The agent experiences one journey. [Agent analytics](https://www.agentrail.co/blog/agent-analytics) is the work of reconstructing that task.

Readiness is the prior question. Were the public surfaces available at all? You can be readable and still lose the task on a bad error body. You can ship a careful MCP description and stay invisible if nothing points at the server. Repair the precondition you failed. Then watch production journeys to see whether the task moves. If you want the metric definitions before you change a page, start with [how to measure agent experience](https://www.agentrail.co/blog/how-to-measure-agent-experience).

## Frequently asked questions

### What is agent readiness?

Agent readiness is whether an AI agent can find a product, read it, use a supported action, and pay when payment is part of the product. The check covers the public website, docs, and app. It uses what an agent can reach without a login and without a private API.

### Is agent readiness the same thing as agent experience?

Agent readiness describes the public surfaces an agent needs. [Agent experience](https://www.agentrail.co/blog/agent-experience) describes what happens when an agent tries to finish a task. A strong scan can sit beside a broken task. A finished task can sit on a site that still fails a discovery check.

### Does a readiness scan need access to our product?

The public scan does not log in, and it does not call a private API. If an agent on the public web cannot see a surface, that surface is outside the score. We do fetch the site with JavaScript and without it, because many agents never run your client bundle.

### Does a high score mean agents complete tasks?

A high score means the public checks passed. Task completion is a different measurement. It asks whether the environment reached the goal state. It does not ask whether a file was present. Use [agent experience metrics](https://www.agentrail.co/blog/how-to-measure-agent-experience) for that rate. Use the [checker](https://www.agentrail.co/agent-readiness/checker) when you want the score itself.

## Start from the public contract

If an agent cannot find the docs, read them, call the tool, or pay, later edits have nothing solid to stand on. See where a site sits on the [agent readiness leaderboard](https://www.agentrail.co/agent-readiness). Then decide whether the gap is a missing surface, or a task that dies after the surface is present. Readiness scans are included with Agentrail. Journeys and simulations are priced on their own on the [pricing page](https://www.agentrail.co/pricing). The company behind the product is on [about](https://www.agentrail.co/about).