---
title: "Why Agent-Readiness Scores Disagree"
description: "We ran three public sites through four agent-readiness scanners, our own included. The same site scored anywhere from 47 to 96 depending on the tool, and every tool ranked the three sites differently. Here is what each one measures, and how to read any readiness score."
author: "Victor Okolie"
publishedAt: "2026-10-23"
category: "AI Readiness"
tags: ["Agent Readiness Score", "Agent Readiness", "Scoring Methodology", "Cloudflare", "Ora", "Developer Documentation"]
canonical: https://glintbase.dev/blog/why-agent-readiness-scores-disagree
---

# Why Agent-Readiness Scores Disagree

> We ran three public sites through four agent-readiness scanners, our own included. The same site scored anywhere from 47 to 96 depending on the tool, and every tool ranked the three sites differently. Here is what each one measures, and how to read any readiness score.

Published 2026-10-23 by Victor Okolie · AI Readiness · HTML version: https://glintbase.dev/blog/why-agent-readiness-scores-disagree

There are now more free agent-readiness scanners than anyone can keep track of. Paste a URL, get a number. The numbers don't agree.

We wanted to see how far apart they are, so we ran three public sites through four scanners, ours included, on the same few days in October 2026.

> **Insight:**
>
> The same site scored up to 43 points apart depending on the tool, and no two tools put the three sites in the same order. A readiness score isn't a measurement of one thing. It's a summary of whichever checks that tool chose, with weights it chose.

## The Results

| Site | ARS 3.3 (Glintbase) | Ora | Cloudflare | Velu (docs) | Spread |
|---|---:|---:|---:|---:|---:|
| glintbase.dev | 86 (A) | 85 (B) | 53 | 59 (F) | 33 |
| docs.stripe.com | 68 (C) | 72 (B) | 47 | 87 (B) | 40 |
| modelcontextprotocol.io | 82 (B) | 69 (C) | 53 | 96 (A) | 43 |

How we got these:

- **ARS 3.3:** our scanner, version `ars-3.3.0`, run on 23 October.
- **Ora:** cached scores from Ora's public score API, scanned between 16 and 21 October.
- **Cloudflare:** isitagentready.com, run on 23 October. It also reports a maturity level, which was level 4 for all three sites.
- **Velu:** the agent-friendly documentation scorer, run on 23 October against `glintbase.dev/docs`, `docs.stripe.com` and `modelcontextprotocol.io`.

Vercel's Is Agentic is built on Ora's checks, so we left it out to avoid counting one rubric twice.

The rankings don't agree either:

- ARS ranks glintbase.dev first, then modelcontextprotocol.io, then Stripe's docs.
- Ora ranks glintbase.dev, then Stripe, then MCP.
- Velu ranks MCP first and glintbase.dev last.
- Cloudflare puts glintbase.dev and MCP level, with Stripe four points behind.

**Two conflicts to be clear about.** glintbase.dev is our own site, and it was built with our own scanner in hand, so it's no surprise it does well on ARS. And ARS's four-layer structure and grade bands come from Ora's published method, which partly explains why our scores sit closest to Ora's.

## Why They Disagree

Each tool answers a different question.

**Cloudflare** checks emerging standards across five categories: discoverability, content, bot access control, API/auth/MCP/skill discovery, and commerce. Its discovery category has eight checks for newer conventions, such as DNS records for agent discovery, Agent Skills and OAuth discovery. All three sites passed the basics but scored one or two out of eight there, which is what pulls every score down to around 50.

**Velu** scores documentation only, with 23 checks across seven categories from the open Agent-Friendly Documentation spec. It crawls docs pages and judges page size, structure, Markdown availability and whether pages point agents to llms.txt. It could discover only one page on `glintbase.dev/docs`, so it left its page-level categories unscored. With fewer categories counted, one failed check pulled the grade down to F.

**Ora** scores four weighted layers (Discovery, Access, Usability and Payments) with 126 checks. It includes whether an agent can operate the product through an API or MCP server.

**ARS** uses the same four layers, but weights them by site type and gives points only for evidence the scanner observed. Checks it can't verify are left out instead of counted as failures. Since 3.3 it also caps a site that blocks agents or hides content behind JavaScript, and it probes MCP servers live. modelcontextprotocol.io scores high on ARS because the scanner found a working MCP server there and verified it, which earns full marks on the usability layer. Stripe's docs score lower because our scanner treats them as a docs site and verified fewer of its checks: 57% of scored points, against 81% for glintbase.dev.

None of these is wrong. A docs-quality score, a standards checklist and an operability score are different measurements that share a 0–100 scale.

## A Different Kind of Tool

All four are checklist scanners: they read what a site serves. A separate group sends a real agent to try a task. Microsoft's Scope, published as an open-source research preview, runs coding agents such as Copilot and Claude Code through a task, scores the result against criteria and keeps the tool-call traces. That measures whether a task got done, which no checklist score can tell you, but it needs a task and a harness for each site.

ARS doesn't measure task success either, and our methodology says so. Whether its checks predict task success is what our next pre-registered study will test.

## How to Read Any Readiness Score

1. **Always name the tool, version and date.** "85 on Ora (scanned 19 October)" and "86 on ARS 3.3 (23 October)" are different claims. A bare "86" isn't.
2. **Read the checks, not the number.** Every tool here lists what passed and failed, and that list is where the value is.
3. **Pick the tool that matches the question.** For docs agents can read, use a docs scorer. For new discovery standards, use a standards checklist. For whether agents can operate your product, use a tool that probes your API and MCP server.
4. **Don't compare scores across tools,** or across versions of one tool. ARS stamps every score with its `score_version` for this reason.
5. **Treat any score as a hypothesis about agent success,** until someone measures task outcomes on your site.

*Scores as shown by each tool on the dates above. Disclosure: Glintbase sells agent-readiness scoring.*

> Check your own site's Agent Readiness Score for free at https://scan.glintbase.dev