A useful website audit should explain what is wrong, what was actually measured, and what remains unknown. Its score should follow that evidence.
Scoring 3.0.0 is our release candidate for a more transparent SEO-GEO Reports methodology. It removes several unjustified penalties, separates measurements that describe different things, and makes missing data explicit. This is a preview of the candidate, not an announcement that production has already been upgraded.
Missing data is not a failed test
A crawler cannot assess content it never received. A timeout, a TLS error or an unusable homepage response should not become an invented accessibility score of zero.
The new model distinguishes an unavailable measurement from a measured failure. An unscored category has a null score and grade, displayed as N/A. A real zero remains zero when checks were performed and failed.
The combined score averages the pillars that were measured. For example, if four measured pillars each score 80 and the fifth is unmeasured, the average is 80. Treating the missing pillar as zero would incorrectly produce 64. This is an arithmetic illustration, not a result from a customer report.
A partial crawl also carries coverage information. Results for usable pages remain useful, but unavailable child pages are not treated as documents missing a title or heading. Reaching the scan limit does not mean the entire website was examined.
Optional content is not a universal requirement
A valid Polish, French or German website does not need an English translation. A monolingual site does not automatically need hreflang. Twitter Cards are optional platform metadata.
The same distinction applies to GEO. A landing page does not universally need an author biography, credentials, sameAs links, citations, an FAQ, a HowTo, a glossary or an llms.txt file. Nor does every page need tables, bullet lists, a summary or question-shaped headings.
The absence of those optional GEO signals no longer creates a scoring penalty, and adding them does not automatically earn a bonus. Useful observations can remain in the report as diagnostic notes. A malformed existing JSON-LD block or an invalid declared date is a different matter: it provides something concrete to validate.
This distinction does not mean all optional markup is useless. It means its usefulness depends on the content and its audience, rather than a universal points checklist.
A GEO score describes a limited technical scope
Authority and Citations currently have no scored checks. Freshness is only evaluated when its conditional date validation applies. When the only scored category is AI Readability, the GEO result describes that narrower set of technical checks.
A GEO score of 100 in that situation does not mean the site has perfect authority, that its claims are true, or that an assistant will cite it. Unscored categories stay outside the weighted denominator; they do not receive free points.
This release does not measure citation frequency or actual visibility in ChatGPT, Claude, Gemini or Perplexity. It also does not turn E-E-A-T into a technical score.
Blocking training is different from blocking search
The robots.txt analysis now evaluates the effective policy for the URL and bot being checked, including Allow exceptions and the most specific matching rule.
GPTBot and ClaudeBot are separated from search-oriented bots such as OAI-SearchBot and Claude-SearchBot. A training-only block does not lower GEO merely because the owner chose to restrict training access.
Google-Extended is treated as a product control, not as a separate Google Search crawler. User-triggered fetchers also have different limitations; robots.txt alone cannot establish how every interaction will behave.
Confirmed search-crawler restrictions remain visible. Allowing a bot in robots.txt still does not guarantee that it will fetch, index or cite the page. A policy check is not a live visibility test.
One scored source per performance dimension
Lighthouse laboratory data, CrUX URL data, CrUX origin data and the crawler's HTTP timing answer different questions.
For LCP and CLS, the scoring prefers a suitable CrUX URL measurement when available and otherwise uses the laboratory fallback. Both observations can remain diagnostic, but they do not become two independent penalties for the same dimension. Origin-level CrUX remains context rather than a substitute result for the requested page.
Total Blocking Time stays separate from INP. TBT can describe laboratory loading responsiveness when an applicable INP measurement is absent; it is never relabelled as a measured INP. A missing PageSpeed or CrUX sample is not itself evidence of poor performance.
The time our crawler spends fetching HTML is also identified as document-fetch timing. It is not LCP or the browser's full page-load time. A good Lighthouse run does not establish that real users meet field Core Web Vitals thresholds.
Shared findings keep their diagnostics
Several confirmed overlaps now have an explicit scoring owner. Examples include shared SEO/accessibility findings, homepage title and H1 overlaps, and compression or caching observations repeated across SEO and Performance.
The secondary observation can remain useful without adding another weight for the same confirmed defect. Findings are linked by stable identifiers where the evidence supports that relationship, rather than merged solely because their titles match.
This does not mean every similar-looking check is a duplicate. Title presence and a separate title-length heuristic, for example, are different conditions. The release is not a claim that all overlap across the engine has been eliminated.
Severity is separate from arithmetic
A missing H1, viewport or meta description is not automatically an emergency. Likewise, a missing Content-Security-Policy header does not prove an exploitable XSS vulnerability.
Severity expresses the interpretation and priority of a finding. It is not an extra scoring multiplier. A warning can still represent a failed, weighted technical check, while a confirmed high-impact exposure can retain Critical severity.
Accessibility findings also need careful wording. Missing main or navigation landmarks do not automatically prove a WCAG failure, and a decorative image can correctly use an empty alt attribute. Automated checks do not certify WCAG conformance or establish the security of an entire website.
For the planned initial rollout, the browser audit is disabled. The report must therefore disclose that checks requiring it, including applicable contrast and focus checks, were not performed.
What happens to existing reports
Stored reports keep their original scores and methodology versions. We are not recalculating old results or relabelling them as 3.0.0.
After the release is deployed, a new scan will use the new methodology. Its score may rise because an optional or duplicate penalty was removed. It can also fall if a previously passing duplicate weight is removed from a category that still has genuine failures. Neither change alone proves that the website improved or deteriorated.
Compare the methodology, measurement sources and coverage before comparing numbers. The methodology explains the current rules, the sample report illustrates their presentation, and the changelog distinguishes the planned release from historical versions.
The purpose of 3.0.0 is a clearer explanation of the evidence: fix confirmed defects, investigate missing measurements, and choose optional content features for your readers. A score remains a bounded technical assessment, not a promise of rankings, citations, accessibility conformance or security.