Transparent methodology
What the AI visibility score means
An AI visibility score is a weighted, point-in-time summary of whether a public website is accessible, understandable, evidenced, technically sound, and represented accurately in optional sampled web-grounded model-family responses. It is not a search ranking or recommendation guarantee.
What does this tool measure?
It combines eight clearly separated modules: deterministic AI readiness, agent readiness, trust and evidence, Google Lighthouse performance, accessibility, best practices, technical SEO, and optional sampled AI visibility. The deterministic scan reads returned HTML plus robots.txt, sitemap.xml, llms.txt and observable interface signals. PageSpeed uses Google’s official API for mobile and desktop.
The result is presented as one overall score, a signal-strength tier from Invisible to Beacon, a radar of the eight modules, a “model council” view of how sampled model families describe and cite the business, and an ordered climb to 100 of the fixes that earn the most points for the least effort.
What can it not prove?
It cannot prove that an AI system will cite, rank or recommend a business; validate whether testimonials or credentials are true; infer third-party corroboration from the target site; or reproduce consumer ChatGPT, Claude or Gemini products. Results change as sites, indexes, models and providers change.
SEO, AEO, GEO and agent readiness
SEO improves discoverability and interpretation by traditional search engines. AEO structures factual, direct answers for answer engines. GEO improves the clarity and evidence that generative systems can retrieve and synthesize. Agent readiness asks whether software agents can access server-rendered content, understand controls, recover from errors and discover advertised machine interfaces. These overlap, but none substitutes for the others.
Weights, non-linear scoring and unavailable modules
Every applicable check shows earned points, available points and a percentage. Not-applicable checks are excluded. Module weights are: AI readiness 25%, agent readiness 15%, trust/evidence 15%, performance 15%, accessibility 7.5%, best practices 7.5%, technical SEO 5%, and sampled AI visibility 10%.
The available weighted result is transformed using 100 × (raw score ÷ 100)². This deliberately makes high scores difficult to earn: technical hygiene and basic crawlability are prerequisites, not proof that AI recommends the business.
The result is also capped by the weakest available model-family maturity stage: not reliably found (maximum 24), found but misunderstood (44), understood but weakly supported (64), trusted but not recommended (79), recommended in limited samples (89), or consistently recommendable (100). A consistent recommendation requires at least two available model-family samples. These stages are point-in-time observations, not a guarantee of future recommendations.
Coverage always states how many of eight modules are available. Missing supporting modules are excluded rather than scored as zero. However, sampled model visibility is essential: if no model-family answer is available, the tool withholds the AI visibility score and shows a separate non-linear website-readiness result instead. This prevents an incomplete scan from looking stronger through weight renormalisation.
PageSpeed lab data vs CrUX field data
Lighthouse lab scores are controlled simulations returned for mobile and desktop. They include performance, accessibility, best-practices and SEO categories plus FCP, LCP, Speed Index, Total Blocking Time and CLS. CrUX is aggregated real-user field data and may be absent when a URL or origin lacks sufficient eligible traffic. We display these separately and never label the scanner’s basic server fetch time as PageSpeed.
Sampled model methodology and reproducibility
Claude first runs three web-grounded buyer scenarios without receiving the company name or domain: a broad category search, an ideal-customer situation and a preference-based situation. The target's unaided recommendation score is the percentage of those three answers that mention the business name, known domain or a returned citation to that domain: 0, 33, 67 or 100. Only after those mentions are counted does a named follow-up diagnose why the business appeared or did not. After the user requests the report by email, OpenAI- and Google-family checks provide additional diagnostic context but do not alter that primary unaided score. The named follow-up also does not retroactively count as visibility. Requested/returned model IDs, timestamp, prompt version, citations and limitations are retained. Provider routing, web indexes and outputs can vary, so these are sampled point-in-time indications rather than consumer-app results.
The original GEO research formalised visibility as a measurable outcome in generated answers and reported that optimisation effects vary by domain. For that reason, this tool separates technical readiness from actual sampled mentions and citations instead of presenting one opaque “AI score.”
Privacy and data handling
The scanner requests only public web resources. URLs may appear in shareable browser query strings and service logs. Deterministic, PageSpeed and unaided Claude evidence is shown before email. An email is requested to run and reveal additional OpenAI and Gemini checks and send the Implementation Pack. Marketing guidance requires a separate optional consent; unsubscribe anytime. The configured CRM and delivery providers own report delivery and follow-up. Do not submit private or credential-bearing URLs.
Who operates AI Visibility Meter?
AI Visibility Meter is operated by Excelsior AI, an Australian AI implementation business. The operator relationship is disclosed throughout the product and in structured data.