Methodology
How Free GEO Checker measures AI visibility
The score combines two things: whether AI assistants actually find and understand a website today, and whether the site is technically ready to be read by them. Everything below describes how the tool works today.
The score
| Part | Weight |
|---|---|
| AI visibility How each tested assistant answers questions about the domain. | 60 % |
| Technical readiness 12 weighted checks on the public homepage and site files. | 40 % |
AI visibility
Each assistant gets two questions. The first lets it search the live web:
Look up example.com on the web. What is this website, and which company or product is behind it? Answer in at most three short sentences and cite your sources.
The second asks it to answer without web search:
From your own knowledge only, without searching: what is example.com? Answer in at most two sentences. If you do not know it, reply exactly: UNKNOWN
The answers are scored per assistant:
| Signal | Weight |
|---|---|
| Found by search At least one search result or citation points to the domain itself. | 40 % |
| Describes you The search answer is graded against a one-sentence summary of what the homepage says the business does: matches (full), partly matches (half), contradicts or no answer (none). This measures agreement with the homepage, not independent fact-checking. | 30 % |
| Answers without search The no-search answer is graded the same way. It shows whether the model can describe the site without searching; it does not reveal where that knowledge came from. | 30 % |
The AI visibility score is the average across the assistants that answered. Grading and the three suggested fixes are produced by a separate, smaller model (Claude Haiku) that compares each answer with a one-sentence summary of the homepage.
Assistants tested today: Claude, through the Anthropic API (model claude-sonnet-5-5 with Anthropic's web search tool; grading by claude-haiku-5-5). Results from the API can differ from the Claude apps. ChatGPT, Gemini, Perplexity and others are added as their integrations go live; the homepage always lists the assistants currently tested.
Technical readiness
The checker reads the public homepage, robots.txt, llms.txt and sitemap.xml, exactly as a crawler would.
| Check | Points |
|---|---|
| AI crawlers allowed in robots.txt robots.txt does not block the bots that fetch pages for AI search answers: OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot and Perplexity-User. Partial credit per bot. | 10 |
| AI crawler user agents not refused The homepage is requested from our server with the user agents of OAI-SearchBot, ClaudeBot, PerplexityBot and GPTBot. A refusal or bot challenge while a normal request succeeds counts as blocked. This detects user-agent rules; it does not confirm access from the crawlers’ own networks. Partial credit per bot. | 10 |
| Content visible without JavaScript Words of visible text in the server-rendered HTML. Full credit at 300 words or more, partial credit below that. | 15 |
| Structured data (schema.org) JSON-LD describing the organisation (Organization, LocalBusiness or a subtype) gives full credit; other JSON-LD gives half. | 15 |
| Indexable and snippet-eligible No noindex in the robots meta tag or X-Robots-Tag header (5), snippets not blocked with nosnippet or max-snippet:0 (2.5), and no canonical tag pointing to another domain (2.5). | 10 |
| llms.txt (optional) A plain-text /llms.txt file exists. Shown as a suggestion without points: some AI tools read it, and Google states it does not affect Search. | 0 |
| Title and meta description A page title (5) and a meta description of at least 50 characters (5). | 10 |
| Heading structure Exactly one H1 (2.5) and at least two H2 subheadings (2.5). | 5 |
| Question-style headings / FAQ At least two headings phrased as questions, or FAQPage markup. | 5 |
| Sitemap A valid /sitemap.xml or a Sitemap line in robots.txt. | 5 |
| HTTPS The homepage is served over HTTPS. | 5 |
| Language declared The html element has a lang attribute. | 5 |
| Open Graph tags An og:title tag is present. | 5 |
Results, links and caching
- Every completed check is saved as a report with its own link, so a shared link always shows the result that was actually measured, with the date and time of the check.
- A domain checked again on the same day returns the saved result, labelled “saved result from” with its original time. A new day gives a fresh check.
- To keep the service free, each visitor can run five AI checks per hour. Technical checks are not limited.
Limitations
- AI answers vary between runs, models and dates. Each scan is a dated observation of how assistants answered at that moment. In our repeatability pilot a site’s AI visibility score had an average standard deviation of about 8 points across identical runs.
- Each question is asked once per check. GEO Platform Pro™ audits repeat every question several times and across markets.
- Only the homepage is analysed technically. Other page templates can behave differently.
- Grading by a language model can occasionally misjudge an answer; the full answers are shown in every report so you can check them.
- A high score reflects good conditions for being found. It does not guarantee mentions, citations or rankings.
Methodology v0.16 · updated 19 October 2026