How we measure AI visibility
The AI Readiness Score (0–100) is built from 33 deterministic technical tests plus live checks in ChatGPT, Perplexity and Gemini. This page explains what we measure, why, and how we keep the method honest. Current version: v0.3, last updated 15 July 2026.
Four principles
Deterministic, not opinion
Every automated test is binary or threshold-based: same website, same test, same result — every time. No test relies on a language model's opinion.
Sourced
Each test cites its source: official crawler documentation from OpenAI, Anthropic, Google, Perplexity and Bing, or open web standards (RFC 9309, schema.org). When docs are silent, we say so and test empirically.
Verified in live engines
Technical readiness is not the goal — being recommended is. We ask real questions in ChatGPT, Perplexity and Gemini and document every answer with screenshots.
Versioned
The AI ecosystem changes fast, so the methodology has a version number and a changelog. Every report states its version, and we never compare scores across versions without labelling them.
What we measure — seven categories
| Category | Weight | Why it matters |
|---|---|---|
| A. Accessibility for AI crawlers | 26 pts | robots.txt rules and firewall behaviour for the training, search and user-fetch bots of OpenAI, Anthropic, Google, Perplexity and Bing. A site blocked at the source is invisible — no matter how good the content is. |
| B. Content without JavaScript | 15 pts | AI crawlers do not execute JavaScript. A single-page app without server-side rendering is an empty page to them. We compare the raw HTML with the rendered page. |
| C. Structured data | 15 pts | JSON-LD (Organization, FAQ, breadcrumbs) is the cheapest way to tell AI unambiguously who you are, what you do and where. |
| D. Content structure | 15 pts | AI extracts answers from clear question-and-answer structure: one H1, clean heading hierarchy, descriptive headings, FAQ sections, proper titles and descriptions. |
| E. Entity clarity | 10 pts | Can a model answer "who, what, where, how to get in touch" without guessing? Contact details, physical address, organisation number, consistent company name. |
| F. Technical foundation | 12 pts | sitemap.xml, canonical URLs, HTTPS, response time, hreflang for multilingual sites — plus a bonus point for the emerging llms.txt standard. |
| G. Actual AI visibility | 8 pts | The proof: does AI know your business, recommend it in your category, and cite your domain? Tested live in three engines; a check passes at 2 of 3. |
Check maximums add up to 101 points (the llms.txt bonus sits on top of the 100 base); the score is normalised: round(100 × earned / 101). Three checks award partial credit — the normalisation and every partial rule are documented, so the numbers in a report always reconcile.
Example checks
A sample of what the 33 tests actually verify — each with its source and a pass/fail criterion:
- robots.txt exists and parses correctly — RFC 9309
- GPTBot is allowed — OpenAI's model-training crawler — developers.openai.com
- OAI-SearchBot is allowed — powers ChatGPT Search citations — developers.openai.com
- ClaudeBot and Claude-SearchBot are allowed — Anthropic's training and search crawlers — support.claude.com
- PerplexityBot is allowed — feeds Perplexity's search index (it does not train models) — docs.perplexity.ai
- The firewall does not block AI user agents — tested empirically with four different bot identities, scored proportionally — repeatable HTTP test
- Meaningful text is present in the raw HTML, before any JavaScript runs — empirical render comparison
- Valid Organization JSON-LD with name and address — schema.org, Google structured-data docs
- Server responds in under 1.5 seconds — threshold-scored TTFB measurement
- llms.txt is present — llmstxt.org (bonus point)
The full list of 33 checks — with per-check scoring, thresholds and sources — is part of the paid audit report, together with your documented results for every single one.
Live tests: what AI actually answers
Category G is the only part measured by hand — deliberately. Three fixed questions per business, always the same engines in the same order: ChatGPT → Perplexity → Gemini, on free tiers (what a real customer uses). Does AI know the business? Does it recommend it when asked for your service in your city? Does it cite your domain? Every answer is screenshotted and attached to the report as evidence, and in monthly monitoring the exact same questions are repeated verbatim — so change over time is measured, not felt.
Reading the score
| Score | What it means |
|---|---|
| 0–39 | Invisible to AI |
| 40–59 | Partially visible |
| 60–79 | Good foundation, gaps to close |
| 80–100 | AI-ready |
A score alone says little — context does. Every report compares your result with the median and the leader of your industry, from our own benchmark:
Benchmark computed with methodology v0.2; cross-version comparisons are always labelled in reports. Full distribution, per-category statistics and the checks businesses fail most: see the benchmark report.
Keeping the method honest
In one year the bot ecosystem went from a handful of crawlers to a dozen with distinct roles — training, search, user-fetch. A methodology that stands still is wrong within months. So: every month we review the official bot documentation of OpenAI, Anthropic, Google, Perplexity and Bing; every quarter we re-validate all checks, thresholds and weights against the state of the market — and any scoring change ships as a new version together with a benchmark re-run, so medians and client scores always speak the same language. And when something big happens — a new crawler appears, an engine changes how it reads robots.txt, a new AI search mode launches — we don't wait for the cycle: the review happens within days.
Example from the current changelog (v0.3, July 2026): the firewall test was extended from one bot to four, after a real audit showed a WAF blocking OpenAI's training crawler while letting search bots through — "blocked for AI" turned out to be four different questions, not one. Claude-SearchBot was added as a new check, and Gemini replaced Copilot in the live-test protocol.
See where you stand
The free mini-audit runs 5 of the 33 checks on your website and shows your preliminary AI Readiness Score in under a minute.
Get a free mini-audit →