Methodology
Every site in the index was processed by the same engine under the same constraints: passive public evidence only, one page per site, no authentication, no privileged access, and no interaction with advertised business tools. Scoring is applicability-aware, so a media site is not punished for lacking an API, and deterministic, so rerunning the same evidence produces the same numbers. The reports are HMAC-signed and publicly verifiable at /verify.
Headline numbers
Across the index the average score is 59 out of 100 and the median is 57. The top score belongs to planetscale.com at 91. At the other end sit some of the most visited properties on the web: google.com at 37, oracle.com at 39, and binance.com at 24. A median in the high fifties means the typical major site is discoverable and readable, but offers an agent almost nothing it can reliably call or verify.
The pattern: infrastructure scores high
Developer-tool and infrastructure companies dominate the top of the table because their product is an interface. They publish OpenAPI documents, typed SDKs, machine-readable pricing, status pages, and precise documentation as a matter of course, since their customers are machines and the engineers who operate them. PlanetScale's 91 is what that discipline looks like when measured.
The other pattern: consumer and finance score low
Large consumer platforms and finance sites cluster at the bottom. Their real interfaces sit behind accounts, session auth, and app stores; their robots directives are restrictive; and their machine-facing documentation is sparse or marketing-led. Google's 37 is a useful reminder that size and engineering talent do not automatically produce agent-readiness — the surfaces an agent needs are simply not the surfaces these businesses expose.
The most common failures
Three gaps recur across the index. First, llms.txt is missing on the large majority of audited sites. Second, almost none publish a machine-readable description of their capabilities, leaving agents to scrape and guess. Third, security headers are weak or absent far more often than their cost would justify. All three are inexpensive fixes relative to their effect on the score.
What the top scorers do differently
The leaders share habits rather than secrets: a concise llms.txt that points to canonical documentation, a valid public OpenAPI document, explicit and consistent pricing, strong security headers, fast pages, and claims that survive comparison with observed behavior. Nothing in the top tier requires novel technology. It requires treating machine consumers as first-class users and keeping every surface consistent with the others.
Honest limitations
This is single-page lab data, not a certification. Each score reflects what one deterministic engine could observe passively on one day, from one entry point, and a site can change its evidence at any time. The index is published as a public leaderboard at /reports with the full per-site reports behind it, so readers can inspect the findings and the exact evidence rather than taking the ranking on faith.