Agent Readiness Compare

Editorial policy

How tools are scored, how facts are sourced, and how errors are handled.

In short

Scanners are ranked on ten criteria, each point from evidence with a source URL and a read date. Vendor claims are labelled as claims, and primary sources come before secondary ones. Errors are corrected in public and never backdated. Scanners are read through public endpoints only, and none is included if reading it would break its terms.

Scores follow a published rubric

Ten scanners are ranked on ten criteria worth 100 points. Each point rests on a source with a read date; a part with no source scores 0, and a part that scored 0 for "not found" is read again on the vendor's own pages before the ranking is published. The seven-criterion reading of 4 November, with scores from 1 to 5, is kept as the earlier record. The method page holds the full rubric and the evidence page every cell.

The ten criteria of the scanner rubric
CriterionPointsWhat we ask
Journey15Do the checks that count toward the score cover finding, reading, acting and paying?
Checks15How many checks does the vendor publish?
Real agents15Does it send a real agent at the site, and on what terms?
Open method10Are the check list, the weights, the scale and dated changes public?
Standards5Do the checks cite the open spec they test, and does the vendor maintain a standard?
Dev fit10A public API, an MCP server, a command-line client, a CI gate.
Peers10A ranked leaderboard or a dated corpus report, and its size.
Fixes10Guidance per failed check, one brief for a coding agent, generated fix files.
Repeat scans5Repeat-scan or determinism data, or at least a stated claim.
Free5Free with no sign-up, and a stated cap.

Platforms that fix or monitor, rather than scan, are profiled with the same sourcing rules but are not scored. AI visibility platforms are profiled for contrast and are not ranked.

Placement and scores are set by the evidence

Editorial rule: a score, or the position of a page, changes only when the evidence changes. Where a profiled vendor has a relationship to Ora, the profile page says so.

Every fact has a source and a read date

  • Each fact carries the URL of the page it came from and the date we read it.
  • Primary sources come first: the vendor's or standards body's own page, specification or API response.
  • A claim a vendor makes about itself is labelled as a vendor claim. We do not repeat it as a finding.
  • A secondary source is used only when no primary source was found, and it is labelled as secondary.
  • Ora's lab findings are written as Ora's lab found, with the date. They are never written as what studies show.

Where two primary sources disagree, the page shows both and says so. See the corrections page for an example.

Corrections are public and dates are never moved

When we find or are told of an error, we check it against the primary source, fix the page and list the change on the corrections page. We do not backdate. A page keeps the date it was first published, and a change is dated when it is made.

The edition date is the day the whole dataset was read. The first edition is 2026-11-04. Each page also shows when it was last reviewed. The Concordance is re-run monthly.

The Concordance reads scanners through public endpoints

The Concordance looks up sixty domains in each scanner on one day, with a pause of two to three seconds between requests and no keys or sign-ups. The Concordance method gives the full detail. The table shows how each scanner was read in the 2026-11-04 edition.

How each scanner was read, edition 2026-11-04
ScannerHow it was readLimit
OraIts public score endpoint by domain; a fresh scan only when no cached score existed, capped at 25 fresh scans for the dayWhen rate-limited, the same-day shared crawl and then the leaderboard snapshot were used, and each cell records which
Cloudflare isitagentready.comIts public scan endpoint on the default check profile, with no tokenThe URL Scanner route needs an API token, which we did not use
is-agentic.comIts public read-only report endpoint, which never starts a scanIts llms.txt says automated bulk scanning is not permitted, so domains without a stored report stay unscored
agent-ready.devIts public badge endpointThe badge does not show a scan date
SiteSpeakAI scannerNone usableIt offers only a web form with a session token, and we do not automate web forms, so every cell is unscored
Mintlify Agent ScoreIts public score pages for documentation sitesA cell counts only when the scored host belongs to the sampled domain

A scanner is included only if we can read it within its terms

We never start scans that a scanner's terms forbid. If a scanner allows no automated lookup, its cells are marked unscored with the reason rather than filled by another route. Cached results can be older than the day we fetched them, so the scan date is kept where the scanner shows one.

The sample's bias is stated

Fifty of the sixty domains come from Ora's public leaderboard, which lists sites that someone scanned with Ora. That favours Ora and is-agentic.com coverage. The other ten are mid-sized business sites that nobody tuned for these scanners, and they act as the control. is-agentic.com is left out of the disagreement measure because Ora runs and scores its checks, so counting it would count one method twice.

Sources

  1. Ora, public leaderboard (source of 50 sampled domains). https://ora.ai/leaderboard, read 2026-11-04.
  2. Ora, methodology. https://ora.ai/methodology, read 2026-11-04.
  3. Cloudflare, isitagentready.com llms.txt. https://isitagentready.com/llms.txt, read 2026-11-04.
  4. is-agentic.com, llms.txt. https://is-agentic.com/llms.txt, read 2026-11-04.
  5. agent-ready.dev, llms-full.txt. https://agent-ready.dev/llms-full.txt, read 2026-11-04.
  6. SiteSpeakAI, AI Agent Readiness Scanner. https://sitespeak.ai/tools/ai-agent-readiness-scanner, read 2026-11-04.
  7. Mintlify Agent Score. https://www.mintlify.com/score, read 2026-11-04.

Last reviewed 2026-11-05