ARS 3.3 is the Agent Readiness Score release that acts on what the October research showed. In how ARS compares with AgentReady, Ora and the IETF draft we listed three things our scoring got wrong. This release fixes them and adds the checks that the MCP changes of the last month called for.
Before shipping, we scanned all 103 companies on our leaderboard with the release code. This post covers what changed and what it did to the scores.
A site an agent can't read now can't score above 47, however much else it gets right. llms.txt no longer earns points for merely existing. And the scanner now speaks the MCP protocol revision that most agent tooling is moving to.
What Changed
1. Gates. Two failures now cap the score at 47, the top of grade D:
- the homepage serves a bot challenge or an HTTP 403 to the scanner (
anti-crawler-challenge-free); - the homepage has almost no text without JavaScript and relies on scripts to render (
content-no-js).
These are the two changes that stopped retrieval outright in AgentReady.org's controlled experiment. Weighted points shouldn't be able to make up for a site an agent can't read. The value 47 is a policy choice, and we say so in the methodology.
2. "Not observed." When the homepage is a challenge or a 403 page, the checks that would read its content are now skipped instead of scored against the refusal page. The block itself is counted once, by the gate. The term comes from the IETF individual draft on agent-readiness evaluation.
3. llms.txt is optional. A missing llms.txt no longer costs points. When one exists it must be a real index: not an HTML error page served with a 200, and with links that resolve. Linking it from your pages earns a bonus, because that's how agents actually reach it. Our llms.txt post has the evidence.
4. MCP, current revision. The live probe now sends server/discover (MCP 2026-07-28) first and falls back to the initialize handshake for older servers, the same procedure as our study of 3,000 MCP servers.
New bonus checks:
- speaks 2026-07-28;
- a Server Card listed in
/.well-known/ai-catalog.jsonor served at the endpoint's/server-card, per the final SEP-2127; - a manageable tool menu (at most 40 tools and about 10,000 tokens of definitions);
- schemas in JSON Schema 2020-12, MCP's default;
- an auth challenge that names a scope.
Logging, Sampling and Roots no longer earn points: the 2026-07-28 spec deprecates them.
5. Better discovery. OpenAPI specs are now found through Link headers, /.well-known/api-catalog and links on the homepage, not only at guessed paths. YAML specs are parsed instead of skipped.
6. Measurement kinds. Every check now says whether its verdict came from a live response, parsed content, mere presence, or local code. A new runtime score counts only live and content evidence, so a site can't score there just because an MCP endpoint or spec file exists.
7. Two reported, unscored checks. poppy-discovery validates a Personal Agent Protocol /.well-known/poppy.json, which is still a draft. robots-edge-consistency flags sites whose robots.txt allows AI agents while their edge refuses the scanner.
The registry now has 137 checks. Scores from ARS 3.2 can't be compared with 3.3: every score carries its score_version.
What It Did to the Leaderboard
We scanned the 103 leaderboard companies with the release code. One, fern.com, didn't respond and wasn't scored. For the other 102:
| ARS 3.2 | ARS 3.3 | |
|---|---|---|
| Average score | 51.5 | 49.1 |
| Grade A | — | 2 |
| Grade B | — | 23 |
| Grade C | — | 17 |
| Grade D | — | 45 |
| Grade F | — | 15 |
39 scores went up, 53 went down and 10 didn't change. Changes ranged from −34 to +25.
Gates did most of the damage. 28 sites failed a gate:
- 17 served a bot challenge or a 403 to the scanner;
- 11 had a homepage that needs JavaScript.
For 18 of them the cap actually lowered the score, by 12.6 points on average. The other 10 already scored 47 or less. All six of the biggest drops, from between 69 and 81 down to 47, came from a gate.
A caution on the bot-challenge gate: the scanner requests pages from one location with a browser user agent. Whether an edge challenges a client depends on its IP address and identity, so a site that challenged us might serve you, and the reverse. In our Silent Wall study, 16–19% of popular sites challenged or blocked our test clients, whether a declared agent, a script or a headless browser, so 17 of 102 is in line with that.
What the release found across the 102 sites:
- llms.txt: 54 serve a real llms.txt. In 46 of those, at least 80% of the sampled links resolve. Only 7 link to it from their homepage.
- Markdown: 30 return Markdown when asked for it, and 7 advertise a Markdown alternate in their page head.
- MCP: we found an MCP server for 57. Of the 33 whose handshake we could complete, 23 speak 2026-07-28. That's far more than the 13% we saw across the whole MCP Registry, which fits a leaderboard full of developer-tool companies.
- Server Cards: 9 have a valid Server Card where the final spec puts it, and 8 more have one in the wrong place or missing required fields.
- API catalog: 9 publish a valid RFC 9727 API catalog.
- Personal Agent Protocol: none publishes a
poppy.jsonyet. - Policy vs. edge: for 13, robots.txt allows AI agents while the edge refused our scanner.
A Bug We Found on the Way
The first full run turned up a bug that predates this release. When checking whether llms.txt links resolve, the scanner stops reading any page over 200 KB, and it counted those pages as broken links even though they had answered with HTTP 200. Large docs pages are common, so many sites failed that check unfairly, in ARS 3.2 as well. It's fixed in 3.3: a page that answers with a 2xx status counts as resolved, however large. The same run showed the JavaScript gate also caught short static pages, so it now applies only to pages that rely on scripts to render.
Also in This Release
- Our own MCP server answers
server/discoverand still serves older clients, publishes a Server Card, and lists it in our AI Catalog. - Packages:
@glintbase/scanner-core,@glintbase/mcpand@glintbase/climove to 3.5.0, on MCP TypeScript SDK 1.32.1.
What hasn't changed: the layer weights, and the caveat that ARS isn't yet validated against agent task success. Our next pre-registered study will test whether ARS's measured checks predict whether agents succeed.
Disclosure: Glintbase sells agent-readiness scoring, and glintbase.dev is one of the scored sites.
Run a Free AI Readiness Audit
Identify broken parameter signatures, missing Model Context Protocol schemas, and sitemap issues in your developer documentation.
