Measured signals, clearly labelled limits.

Rubric v6 scores 48 checks across five axes. One check measures the sitemap and a bounded sample of its URLs; separate sample diagnostics show missing page metadata and repeated titles without changing the LR weights. This keeps ratings comparable when a site has no sitemap or only one URL. The number in parentheses is each check’s relative weight inside its axis; the five axis totals determine the final 100 points. Every check has its own page in the audit check library, stating what it measures and where it can be wrong. v4 added AI visibility and re-took the other four to pay for it; v5 split the crawler check by consequence and added three security checks; v6 measured what the competitors’ single-URL audits measure and this one did not — canonical destinations, soft 404s, URL shape, rich-result types, dates, compression, asset caching and image markup. A report keeps the axes it was scored on, and the version is printed on every report. Rubric last updated 29 October 2026.

• Discoverability · 30 points

  Title tag (12), meta description (10), H1 heading (7), canonical URL (3), canonical target (4), indexability (3), URL structure (2), unknown URLs 404 (3), language attribute (2), favicon (3), sitemap.xml (3), sitemap URL quality (4), robots.txt (2). The canonical target check reads where the tag points, because a canonical aimed at another page is the most expensive silent failure this axis can find. The 404 check asks for a random path: a status of 200 means every typo renders a real page. The sitemap quality check can include failures found in sampled pages.

• Shareability · 18 points

  og:title (3), og:description (3), og:image (6), twitter:card (4), twitter:image (2). Since v4 the og:image URL is actually fetched: a declared image that returns 404, serves HTML, or weighs megabytes no longer passes.

• AI visibility · 15 points

  Answer-engine access (6), server-rendered content (4), llms.txt (3), structured data entity (3), quotable structure (2), training-crawler consent (1, descriptive). v5 split the crawler check because the two groups cost different things: OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-WebBot and anthropic-ai decide whether a live answer can cite you today, while GPTBot, ClaudeBot and Google-Extended decide whether your words enter a future model. The first group is scored; the second is reported and never marked down, because refusing training data is a decision, not a defect. Access is parsed per User-agent group, longest match wins, an unnamed agent falls back to the * group. Server-rendered content fails only when the HTML looks like a client-side app shell — an empty root plus a “enable JavaScript” noscript. Quotable structure counts subheadings, lists and mean paragraph length, because an answer engine quotes a passage rather than a page. llms.txt is scored as an emerging convention: absent costs less than a broken share image does, and there is no evidence it moves Google rankings.

• Credibility · 22 points

  Image alt text (8), heading structure (6), body copy (7), E-E-A-T trust signals (4), experience signals (3), expertise signals (3), authority signals (3), JSON-LD syntax (4), rich-result markup (3), published or updated date (2), internal links (4), mixed content (4), response headers (3), form transport (3). Missing JSON-LD is optional and keeps 40% of the markup check rather than failing outright; malformed JSON-LD fails. The date check reads only machine-readable dates — article meta tags, <time datetime>, or JSON-LD — because a visible “updated last week” is prose, not data. The three security checks landed here rather than in Performance because they describe whether a visitor’s browser can trust the origin, and headers are scored on presence, not quality.

• Performance · 15 points

  Server response time (8), text compression (3), static asset caching (3), HTML weight (4), blocking scripts (5), mobile viewport (3), image dimensions (3), responsive and modern images (3), HTTPS (2), lazy-loaded images (2). Compression and cache policy are read from the headers of the HTML response and of the first same-origin asset; the image checks read the markup, not the file. This is what a handful of requests from our server can measure — it is not Core Web Vitals, not field data and not a lab score. Run Lighthouse for that.

Checks earn partial credit. Relative weights are normalized within each axis, then the axis result is scaled to its published maximum. This 0–100 audit result is the Audit Launch Rating (Audit LR).