Coverage contract · version 2 · 5 November 2026
Website readiness and observed answers answer different questions. GEOrank keeps them separate.
A model-generated review of a public website across content, authority, technical indicators, AI-readiness indicators and freshness. Checklist totals follow the displayed scoring weights. The audited page's HTML is fetched directly for schema, meta and Open Graph tags, canonical link, robots.txt, sitemap.xml and llms.txt; those checks are labelled measured, and a failed fetch is reported as failed. Other pages listed are model-reported, not independent fetch attestations.
Up to five generated buyer questions, each asked through the official APIs of Claude, OpenAI (the GPT models behind ChatGPT), Gemini, DeepSeek, Perplexity and Grok with live web search, in manually initiated runs. Inspect successful answer text, returned citations, method, model when supplied and capture time for each engine. The GEO score is GEOrank's own custom index, not an industry measure: each answer scores 100 if it cites your site, 50 if it only names your brand and 0 otherwise, and the GEO score averages the engines' scores, each weighted equally. The workspace shows the mention rate, citation rate and each engine's result separately so you can see what produced the index. Suggested questions are hypotheses, not demand data. Each question is sent as a buyer would ask it, starting only with "Search the web and answer in English" and saying who is asking; answer length and the tools named are measured afterwards. Question sets created before 7 November 2026 also asked for 3–4 sentences and no more than 120 words, and keep that wording when repeated so their comparisons stay like-for-like.
Account-owned report snapshots can save and reopen. These client-submitted snapshots are not server-attested provider records. Competitor domains provide context; they are not automatically matched competitor measurements.
A 100-point score calculated only from what GEOrank fetches: the page's raw HTML and response headers, robots.txt, sitemap.xml, llms.txt and llms-full.txt, plus a second fetch of the page as each live AI search crawler. No model is involved, so the same captured inputs and scoring version give the same score. Network responses and crawler access checks can change between fetches, so two fetches of one page can differ. It measures whether AI crawlers can reach and read the page. It does not measure whether AI answers mention you, and the weights are GEOrank's own choices, not an industry standard.
15 for live AI search crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot, Bingbot) allowed in robots.txt and served the normal page; 5 for assistant link fetchers (ChatGPT-User, Claude-User, Perplexity-User); 5 for a sitemap found and declared in robots.txt. Minus 10 if the page says noindex. Training crawlers are reported but not scored.
15 for readable text without running JavaScript (300+ words; 8 for 100–299; 0 for a JavaScript-only page); 5 for one H1 and section headings; 5 for a title, meta description and canonical link.
8 for valid JSON-LD in the raw HTML (4 if some blocks are invalid); 6 for an Organization plus a product, service or business type; 6 for FAQPage, sameAs links and a description or disambiguatingDescription.
8 for /llms.txt served as text, not a web page; 4 for its format (an H1 title, a summary blockquote, markdown links); 3 for /llms-full.txt or an llms.txt link in the page head.
Why it is weighted: llms.txt is a proposed convention for giving AI tools a plain-text map of a site. We have not seen public evidence that AI search engines use it to choose sources, and its presence is not a validated predictor of being recommended. It is scored as a low-cost readability step, and these 15 points should not be read as a likelihood of appearing in answers.
6 for a table or definition list in the raw HTML; 5 for sections averaging 120–350 words under H2/H3 headings (2 for 50–600); 4 for a dateModified, article modified time or Last-Modified header within 180 days. Minus 5 for an 800-word stretch with no heading.
These are not measurements from the consumer interfaces of Claude, ChatGPT, Gemini, DeepSeek, Perplexity, Grok or other engines. A regional database does not establish a regional measurement.
Name mentions and owned-domain citations are separate observations. A mention does not establish a recommendation or sentiment. Rates use successful answers only; failed checks are not zeros. A URL citation uses the exact owned host or its subdomains, not a matching string in an unrelated hostname.
A readiness total is a calculation over model assessments, not observed answer visibility, a statistical confidence interval or a probability of recommendation. Small samples do not support market-share or causal-impact claims. Where available, Google PageSpeed results are separate; unavailable results are not silently replaced with invented figures.
Review and version one to five buyer questions, brand domains, competitors and audience before a manual run. Inspect answers and sources, then create an evidence-linked task with owner, priority, completion criteria and an implementation record. Owners are text assignments, not team invitations or notifications.
Completed tasks can manually repeat their retained cohort. Each task retains one baseline answer, so its before/after comparison pairs only that question. Failures and incompatible models, methods, scope or timing are excluded rather than counted as zero visibility. Date-only implementation records require a baseline before that UTC day and a recheck after it.
The private Actions & Report view provides CSV and browser print/PDF with task evidence, revisions and rechecks. These exports reflect the current page; check save status before closing. Saves are limited to 20 tasks and 3 MB per snapshot. No old evidence is silently pruned to fit.
Generally available unattended monitoring, verified locale execution, sustained competitor trends, causal impact attribution and role-based permissions beyond a team owner and members are not available. Weekly automated monitoring runs as a private pilot for one approved project only. Manual API usage is rate limited, not an unlimited subscription or a guaranteed invoice cap. A recheck attempt is not proof of improvement, traffic or revenue.
Use a mature monitoring product if those capabilities are essential today. GEOrank currently fits a small, inspectable investigation. No ranking, recommendation, traffic or revenue outcome is guaranteed.
Software for founder-led B2B businesses and lean marketing teams. Model-generated findings require your review before implementation. No analyst-review service, response-time commitment or monitoring subscription is offered here.
Contact Stainaz →