How SEO-GEO Scoring and Metrics Work

1. The crawl: how we collect data

Analysis starts with a focused crawl of your domain. After validating the URL (only public http/https hosts are allowed — private, localhost and cloud-metadata addresses are rejected for security), the crawler discovers pages via your XML sitemap (referenced in robots.txt) and by following internal links in the page structure.

• Page budget — free reports crawl up to 300 pages (the engine supports up to 500). Larger sites are sampled, not fully crawled.
• Fetch limits — each request times out after 10 seconds, with a short delay between requests and a dedicated SEO-GEO-Bot/1.0 user agent.
• Structured HTML analysis — from v1.1 onward we parse the returned HTML as a document tree (titles, meta tags, headings, links, images, structured data) rather than scanning the raw markup with text patterns. This improves accuracy; we still do not execute JavaScript, so content injected only client-side is not measured.
• Site files — we additionally fetch robots.txt and llms.txt, and record per-page load time, TTFB, size, internal/external links, broken links and orphan pages.

See how it works for the end-to-end flow from URL to report.

2. The scoring model

Each analysis runs independent category analyzers across four pillars — four categories each for SEO, GEO, WCAG and Security. Every analyzer runs a fixed set of pass/fail checks and produces a normalized score:

Each check has a weight (usually 1; critical signals such as HTTPS, indexability, author identity or LCP use weight 2). The passedChecks / totalChecks counts in the report and below still refer to the number of checks, not their weights. Category scores are then combined with fixed weights:

The combined score is the average of the four pillar scores: round((SEO + GEO + WCAG + Security) / 4). Reports generated before a pillar existed average only the pillars they include. Each 0–100 score maps to a letter grade using the scale below.

3. Grading scale

4. SEO checks in detail

Technical — 30% · 13 checks

Crawlability, indexability, delivery and a mini accessibility audit. Missing HTTPS, a non-200 homepage, a noindex tag, or a missing viewport are weighted as critical (2×) because they can block ranking outright. From v1.3 we also check for main/nav landmarks and form field labels.

Content — 30% · 13 checks

On-page fundamentals across all crawled pages: titles and descriptions (presence, length and uniqueness), heading hierarchy, image alt text, social metadata, and an English-coverage heuristic for international reach. From v1.3 the report also surfaces thin pages (fewer than 300 words) and flags when the declared html lang does not match the detected language of the homepage body — these appear as metrics and issues, not as additional scored checks.

Structure — 20% · 8 checks

Site architecture: URL hygiene, internal linking strength, orphan pages, breadcrumb navigation, structured-data presence and click depth (measured from URL path segments).

Performance — 20% · 7–10 checks

Server-side delivery speed averaged across crawled pages. Load time is the heaviest single check (worth 2 points). When a PageSpeed Insights API key is configured, three additional lab checks (LCP, CLS, INP on mobile) are measured for the homepage via Google Lighthouse — otherwise only crawler timings apply.

5. GEO checks in detail

AI Readability — 35% · 19 checks

How easily generative engines can parse and reuse your content. This category merges 16 on-page readability checks with 3 AI-access checks (robots.txt/llms.txt); structured-data coverage below counts for 2 of those points. From v1.3 we add four GEO signal checks: Article schema on blog URLs, Speakable markup, glossaries/definitions and quantified facts in body text. Blocking AI crawlers is critical: if GPTBot, PerplexityBot, ClaudeBot or Google-Extended are disallowed, your content cannot be cited at all.

Authority (E-E-A-T) — 25% · 12 checks

Experience, Expertise, Authoritativeness and Trust signals that make a source credible enough to cite: author identity and Person schema, organization identity, an entity graph via sameAs, contact and about pages, credentials and publication dates.

Citations — 20% · 7 checks

Evidence density. Generative engines prefer content backed by outbound links, references, attributed quotations and — most strongly — concrete statistics. Statistics are detected via percentages, large numbers with units, and phrases such as "according to" / "research shows".

Freshness — 20% · 7 checks

Recency signals. Content age is derived from the newest valid datePublished/dateModified found — never from a stray date — and mentions of the current year ignore footers and copyright lines, where the year is usually auto-generated.

6. WCAG accessibility checks (v2.0)

From v2.0 the report includes a dedicated WCAG branch aligned with the POUR principles (Perceivable, Operable, Understandable, Robust). It is a static HTML audit — we do not render CSS or execute JavaScript, so color contrast, focus order and keyboard traps require manual or browser-based testing (e.g. axe DevTools).

• Perceivable — alt text on images (WCAG 1.1.1), lang on <html> (3.1.1).
• Operable — main/nav landmarks, skip-to-content link (2.4.1), form labels (1.3.1, 3.3.2), link purpose (2.4.4), button names (4.1.2).
• Understandable — page title (2.4.2), single H1, sequential heading hierarchy (1.3.1).
• Robust — unique IDs (4.1.1), table headers, iframe titles, ARIA on non-semantic elements.

The mini accessibility audit in Technical SEO (landmarks, labels) remains for backward compatibility; the WCAG branch provides the full score and prioritized recommendations.

8. Sub-algorithm: Content Clarity

Content Clarity is a 0–100 readability metric computed on the homepage and blended from three weighted components:

• Paragraphs (35%) — ideal length 50–300 characters; paragraphs over 400 chars lower the score.
• Sentences (40%) — ideal length 40–120 characters; sentences over 180 chars are penalized, extremely short ones slightly so.
• Structure (25%) — rewards at least 2 H2/H3 headings and roughly one heading per ~7 paragraphs.

9. Sub-algorithm: answer-first & question headings

Research on generative engines (Princeton, Aggarwal et al.) shows that AI systems preferentially cite passages that answer the heading's question in the first 40–60 words — so-called heading-to-passage anchoring. We extract every H2/H3 section and its leading paragraph, then:

• · A section passes if its lead is a concise, self-contained answer (roughly 10–80 words, ideal 40–60).
• · The answer-first ratio is the share of sections with such a lead; ≥50% passes the check.
• · Separately, headings phrased as real questions (starting with how/what/why… in EN or PL, or ending with "?") match conversational AI queries and score their own check.

10. Sub-algorithm: AI crawler access

We parse robots.txt into per-user-agent groups (honoring stacked User-agent lines) and check each known AI bot — GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, anthropic-ai, Google-Extended, Applebot-Extended, CCBot, Bytespider and more. A bot counts as blocked when its own group has Disallow: / without an overriding Allow: /. Two failures here are critical (blocked AI bots; a site-wide User-agent: * / Disallow: /), and the presence of an llms.txt index is scored as a positive signal. The report also renders a full allowed/blocked table per engine.

11. Issues, severity and prioritization

Every failed (or partially failed) check produces an issue with a severity — critical, warning or info — and an impact value from 1 to 10. Severity does not directly subtract points (the score comes from the weighted pass ratio); instead it drives prioritization. Recommendations are sorted first by severity (critical → warning → info) and then by descending impact. From v1.3 the HTML report also groups recommendations into Quick wins (lower effort, impact ≥ 5) and Larger projects (higher effort or structural fixes) so you know what to tackle first.

12. Limitations

• Snapshot only — we analyze the site at crawl time. We do not track rankings or historical SERP data.
• No JavaScript rendering — we read server-returned HTML, so content injected only by client-side JS is not seen.
• No backlink or off-page data — scores exclude link authority and external signals; the focus is on-site.
• Crawl depth — free reports crawl up to 300 pages; larger sites are sampled and some pages may not be analyzed.
• Heuristics — signals such as language/English detection, author info and statistics use content heuristics and can occasionally mis-classify.
• GEO evolves — AI citation behavior changes; GEO criteria reflect current best practices and published research and will be updated.
• Passive security only — the Security branch inspects headers, HTML and read-only probes. It does not test server-side logic, authentication or run active penetration testing, so a good score is not a guarantee of security.

For more on what the tool is best used for, see use cases and FAQ.

Credibility and updates

We update this methodology whenever we change scoring or add checks. The "Last updated" date at the top reflects the last substantive change. The same rules run for every URL — there is no paid placement in recommendations. Our team and editorial approach are described on the About page.

Why we publish this (editorial transparency)

We believe audit tools should be transparent about what they measure and how. This page exists so you can interpret your report correctly and compare our approach with other tools (e.g. Ahrefs, SEMrush, Screaming Frog). Scores are calculated from the same rules for every URL. If you have questions, contact us or read our FAQ.