Earn your position. Show the evidence.

The formulas

H = 100 × (1 − broken weight / evaluated weight). Navigation links weigh 3, other homepage links 2, and content links 1. Internal destinations retain the ×2 multiplier. A destination counts once at its most prominent surface in the selected sample. The exact broken-weight ratio decides link-health order.

Q = 0.25 × Structure + 0.25 × Accessibility markup + 0.25 × SEO + 0.10 × Sharing + 0.15 × Security. Each component is a percentage of its applicable checks. A page-level check earns its exact passed/evaluated-applicable proportion. For equal link weights, 1 broken link in 10 and 10 in 100 both give H = 90. Category ratios are combined before rounding. Q is published with two decimal places; the published number is also the quality tiebreak. Equal H and published Q are separated (rank-v5) by the two diagnostic categories that carry no weight in Q: the published AI access category score first, then the published Performance category score (HTML transfer budget). A measured category beats an unmeasured one; two unmeasured categories are equal. Only then does who reached that result first count: the site reigning since the earlier date keeps the position until someone beats it. Only when both dates are equal does the larger verified sample (more evidence, stronger claim) decide. A full tie still shares the position.

AI access and Performance never change Q: blocking AI bots or shipping a heavy page loses no points; those categories only separate sites that are exactly equal on everything scored. HTTP timing remains diagnostic information and never breaks ties, and neither does a random draw or the domain name. Sponsorship, clicks and payment cannot affect a competitive score. Rescans can reveal genuine changes or different measurement availability; this is why coverage and retained evidence are disclosed.

A repeatable selection

We select the homepage and up to 19 same-site pages linked from it, prioritizing navigation surfaces and then normalized URL order. Selection happens before the crawl deadline can shorten the visited set. From these sources we select up to 200 internal destinations by surface weight and URL order. The exact source URLs, destination URLs and weights form a stored sample manifest. A sample with fewer than ten evaluated destinations refreshes on the next scan so a growing project can qualify. Eligible rescans reuse their manifest for 30 days and fetch the URLs again even if their links have disappeared from the HTML. A method change, a new canonical homepage or the 30-day expiry starts a new baseline.

At least ten internal destinations must be verified in the current attempt. Other unknown links, missing source observations or unavailable checklist resources do not disqualify the site. Unknown links are excluded from the evaluated-weight denominator and their coverage is disclosed. A site with fewer destinations still receives a free diagnostic report. Redirect aliases do not give the same final HTML multiple votes. A selected page with a confirmed 404/410 stays in the quality denominator as an unavailable page; it does not disappear and raise the score.

The extended diagnostic report can inspect up to 100 pages, depth two, and discover 1,500 links. These additional pages and external links help find problems but cannot dilute failures in the competitive sample. A changed selection starts a new comparison baseline; its aggregate difference is not presented as a verified improvement.

Confirmation by item

A first pass with at least ten verified internal destinations can qualify. One additional pass rechecks that exact manifest after at least five seconds. Eligibility does not require every item to agree. A newly observed item can use a conclusive reading; a change to fresh previous evidence requires two agreeing observations. Conflicting new readings remain unknown. No best-of-two score is selected.

A network failure is not a missing HTML declaration. A title absent from readable HTML is a scored failure; an unreadable page or unavailable resource is unknown. Each previously measured link or checklist item can retain its prior evidence for up to eight days from its last reliable reading. A failed or conflicting reading does not renew that timestamp. Retained evidence and coverage are disclosed. A subsequent scan excludes expired item evidence; the result itself does not expire by age.

If fewer than ten destinations can be verified now, the latest attempt is inconclusive and does not renew an existing result. A verified result stays ranked until a newer verified result replaces it; the ranking shows its measurement date. After the seven-day report cache anyone can request a rescan. Network restrictions, robots policies, bot walls, timeouts and server errors are unknown evidence, not proof of broken or healthy links.

Each pass retains its 250-second network deadline. An admitted job runs at most two passes; crawler attempts remain bounded to three. Confirmation uses the same domain lock and admission cap. Verification is free. No automatic repeated scan loop is enabled.

What counts as a correction?

A verified repair requires a previously confirmed broken destination to be healthy in the two new observations of the same manifest. Destinations are counted once. Additional coverage or a different baseline does not earn a repair count. A removed URL is not described as a repaired destination. Rank movement is frozen against the same catalog at publication; historical snapshots are preserved.

Open the selected sample and its checks in the report to inspect the evidence. The extended diagnostic checklist is labeled separately because its coverage may differ from the competitive result.

What the checklist can establish

Structure checks headings, main landmarks, viewport, image dimensions and structured-data syntax. Accessibility markup checks language, image alternatives, descriptive link text and names for controls, buttons and links. SEO checks titles, descriptions, self-referencing canonicals and sitemap availability. Sharing checks social metadata and homepage image resources. Security checks HTTPS, insecure embedded resources, and the HSTS and Content-Security-Policy headers of the final homepage response. Not-applicable elements earn no invented points. Confirmed failures lose their applicable points without excluding the site. Unknown observations have no scoring vote when there is no fresh previous evidence; not-applicable checks also have no vote. When a whole category is unavailable, Q uses the published weights proportionally across the available categories. Coverage remains visible, since partial evidence is a limitation, not a claim of complete quality.

This is static HTML evidence, not WCAG certification, source-code review or a measurement of browser interaction. Contrast, keyboard operation and Core Web Vitals are not established by these checks. HTTP response time is not browser loading speed. A bounded sample cannot prove that every page or visitor journey works.

New in quality-v4

• Self-referencing canonical (SEO): exactly one canonical link equal to the page's own final URL; only a trailing slash or fragment may differ. Reference: Google Search Central, canonicalization.
• Image dimensions declared (Structure): every visible image declares width and height or an inline aspect-ratio. Reference: web.dev, layout shift.
• Descriptive link text (Accessibility): no visible link named only "click here", "read more", "saiba mais" and similar. Reference: WCAG 2.4.4 Link Purpose (In Context).
• Structured data syntax now scores under Structure. Reference: Google structured data and schema.org.
• Security (15%): HTTPS and no insecure embedded resources per page, plus two homepage-level observations from the final response headers: Strict-Transport-Security with max-age of at least 15552000 seconds (not applicable on an HTTP homepage) and a Content-Security-Policy header or meta element. Presence is verified, not policy strength. References: OWASP HSTS, OWASP CSP.

Reports measured under quality-v3 keep their recorded checklist until the site is rescanned. The ranking compares only current quality-v4 results; an older checklist is not converted or re-weighted.

Battles and sharing

The competitive winner follows the exact ranking order (H, published Q, AI access, Performance, reigning since, larger verified sample) and requires two qualified contenders. Diagnostic stages provide context; a browser-only result cannot beat a missing standard result. Shared results freeze both measurements and their verdict. New measurements create a rematch. A full tie draws. There is no random, latency or paid tiebreak. A round highlight does not override the overall ranking rule.

Old reports and milestones retain their recorded method. They are not retroactively converted into new measurements. A new standard-v6 scan is required to qualify under this policy.

Limits and references

We validate public destinations and redirects, respect robots.txt, cap response bodies and pace requests. Cached reports are free; signed-in fresh requests retain the 10-per-24-hour quota (admins are exempt; rejected requests and failed scans do not count) and 15-minute per-site interval. A new methodology bypasses the old result cache without bypassing admission limits.

The weights and sample limits are published product rules, not statistical confidence estimates. Reference concepts: W3C page structure, accessibility evaluation limits, and Open Graph.