# What is a good agent readiness score? 5 well-known sites scored 61.8–88.5

> No agent readiness score is good on its own. 5 well-known sites scored 61.8–88.5 in October 2026, and the level says more than the number does.

- Source: /learn/agent-readiness-score-what-is-good
- Updated: 2026-11-01
- Reading time: 7 minutes
- Tags: agent readiness, score, levels, scan records

---
No agent readiness score is good on its own, and none of the three tools compared below states a pass mark. In AgenticWebScore's scan records of 20 October 2026, 5 well-known sites scored from 61.8 to 88.5, and this site's own record scored 94.7. Read the level first. It says what an agent can do on the site, and the number cannot buy or block it.

## What did 5 well-known sites score?

The 5 records below are stored scans of cloudflare.com, resend.com, vercel.com, github.com and developer.mozilla.org. All were made within 1 minute on 20 October 2026 under rubric v0.9. The last row is this site's own record from 25 October 2026.

| Site | Score | Level | Web standards | Agent conventions |
| --- | --- | --- | --- | --- |
| [cloudflare.com](/aws/cloudflare.com) | 88.5 | L5, Agent-native | 92.4 | 79.4 |
| [resend.com](/aws/resend.com) | 86.3 | L5, Agent-native | 98.8 | 57.1 |
| [vercel.com](/aws/vercel.com) | 71.5 | L3, Machine-readable | 83.7 | 42.9 |
| [github.com](/aws/github.com) | 62.8 | L2, Crawlable | 79.0 | 25.0 |
| [developer.mozilla.org](/aws/developer.mozilla.org) | 61.8 | L3, Machine-readable | 88.3 | 0.0 |
| [agenticwebscore.com](/aws/agenticwebscore.com) (this site) | 94.7 | L5, Agent-native | 100.0 | 82.4 |

The median of the 5 external scores is 71.5. Two pairs show why the number alone misleads. developer.mozilla.org (MDN) has the lowest score and holds level 3, while GitHub scores 1.0 higher and holds level 2. Resend and Vercel are 14.8 points apart, which is a gap of 2 levels.

[What 5 well-known sites serve to agents](/learn/five-sites-agent-readiness-october-2026) explains each record check by check.

## How is the score calculated?

The score is 70% of the web-standards part plus 30% of the agent-conventions part. Each part is the share of points earned over the points of the checks that count for the site, from 0 to 100. The [published formula](/methodology#formula) has one more rule: an essential check counts 3 times its weight, and a recommended check counts once.

I recomputed all 6 scores above from their stored rows and the formula. Each matched the stored score to the first decimal. For MDN the sum is 0.7 × 88.3 + 0.3 × 0.0 = 61.8.

That sum has a consequence. A site that uses no agent convention cannot score above 70.0, however good its web standards are. A site that earns every counted point scores exactly 100.

The score is a share of what applies to the site, not of all 60 checks. A check that is out of scope leaves both sides of the fraction. MDN had 14 of its 28 scored web-standards checks out of scope, Vercel 6. So two sites with the same score may have been asked different questions.

## Why does the level matter more than the number?

A level is a set of gates on named checks, and the score never buys one. Level 1 needs a readable homepage. Level 2 adds `robots.txt` and a sitemap that parses. Level 3 adds text, a title, headings and structured data in the HTML, with no JavaScript run.

Levels 4 and 5 apply only to a site where the scan saw an API or agent surface. Level 4 needs a working MCP server or a documented API. Level 5 adds a listed surface, a published status and any OAuth described. The [level table](/methodology#readiness-levels) has the exact checks.

The records show how little the points say about the gates:

- **Vercel, 71.5, level 3.** The record names one check that blocks level 4: no OpenAPI 3.x document answered at the 5 paths the scan requested. On the published formula, moving [openapi](/checks/openapi) from fail to pass, with every other row unchanged, gives 79.9 and opens level 4. The soft 404 in the same record is worth 0.7 points and holds no level.
- **GitHub, 62.8, level 2.** The homepage carries no structured data, so [jsonld](/checks/jsonld) blocks level 3. The same arithmetic gives 71.9 for that one check and level 3.
- **MDN, 61.8, level 3.** The scan saw no API or agent surface, so level 3 is its ceiling, and it reaches it. Of the 38.2 points it lacks, 30 come from the conventions part, which holds no level.

A new scan reads the whole homepage again, so a real result can differ from these what-if sums. The point is the order of work. Fix the check that blocks the next level before the checks that only move the number.

The score is still the finer instrument. Two sites at level 3 differ by 9.7 points in these records. Use it to compare your own site with its earlier records under the same rubric version. A score from v0.8 and one from v0.9 are not comparable, because v0.9 changed how the parts blend.

## How do Cloudflare's isitagentready and Lighthouse report their results?

They report different kinds of number, so a result from one cannot be converted into another. I did not run either tool on the 5 sites. The table compares what each says about its own reporting, from their pages as read on 31 October 2026.

| Aspect | AgenticWebScore | Cloudflare isitagentready | Lighthouse agentic browsing |
| --- | --- | --- | --- |
| Result | A score from 0 to 100 and a level from L0 to L5 | A gauge out of 100 and a level from 0 to 5 with a name, per its page script | A pass ratio. The page says there is no weighted 0–100 score |
| Weights | Published: tier factors 3 and 1, parts 70% and 30% | Not stated on the page or the blog | Not stated on the category page |
| What it requests | Plain HTTP requests, no JavaScript run | Requests to the site, in 4 dimensions that count | A Chrome page load, Chrome 150 or later, experimental |
| Pass mark | None | None stated | None stated |
| Date of the source | Rubric v0.9, 20 October 2026 | Blog of 17 April 2026, updated 15 July 2026 | Page updated 5 May 2026 |

Three details matter when two numbers disagree:

- **What counts.** Cloudflare's [post](https://blog.cloudflare.com/agent-readiness/) says commerce standards do not currently count towards its score, and that by default it checks markdown negotiation but not `llms.txt`. Here, [llms-txt](/checks/llms-txt) is a scored convention.
- **How it is summed.** The script on isitagentready.com divides passed checks by counted checks. Every check weighs the same, and commerce is left out.
- **A check that could not run.** As far as that script shows, it stays in the total. Here, a check that is not measured leaves the score and lowers the confidence instead.

The script also styles a gauge differently below 50 and from 90 up, but the page prints no word for either band. The [comparison page](/compare) lists the other tools, and the [Lighthouse audit-by-audit article](/learn/lighthouse-agentic-browsing-vs-agent-readiness) maps its 7 audits to our checks.

## How big is the sample, and what does it not show?

A sample of 5 sites is small. We chose them: well-known, developer-facing sites with public documentation, scanned in a single minute. They are exemplars, not a benchmark, and the median of 5 says nothing about where the median site on the web sits. We hold no distribution of scores across the web.

The 94.7 comes from a site built against this rubric, by the people who wrote it. Read it as the score of a site tuned to the rubric, not as a target others should expect. It also moves: the previous record of this site, from 20 October 2026, scored 89.5 at level 4.

A score also says only what an agent can find. It does not show that a described endpoint works, with one exception: the [mcp-server](/checks/mcp-server) check opens a session and lists the tools. Each record carries a confidence figure for how much of the rubric the scan could observe, and all 5 external records show 100.

## How should you read your own score?

1. Start with the level, and compare it with the ceiling in the record. The ceiling is 3 for a site with no API or agent surface, and 5 for one with it.
2. Read the failed scored web-standards checks. They are defects, not choices.
3. Read the conventions part as choices. A 0 there is allowed.
4. Check the confidence, and compare only with earlier records of the same site under the same rubric version.

To see the score, the level and the evidence behind each check for your own site, [scan it](/scan).