Source: https://aiexorcist.com/methodology
Last-verified: 2026-10-27

Summary: This mirrors the human page: how public company reports are produced.

Methodology (AIO score + Ghost Test)

This mirrors the human page: how public company reports are produced.

What an audit looks at

Most AI crawlers (GPTBot, ClaudeBot, PerplexityBot) read the HTML a server sends and do not run JavaScript. Each audit samples important pages (found via the XML sitemap and homepage links), fetches each as a plain HTTP request, and also loads it in a real browser. Scores are based on the server HTML. Reports cover that sample, not every URL.

AIO score (0–100)

The sum of four pillars. Page-level pillars are averaged across sampled pages, with pricing and docs pages weighted most.

• Technical & crawlability (25) — robots.txt allows GPTBot, OAI-SearchBot, Claude-SearchBot and PerplexityBot (up to 10); XML sitemap reachable (10); llms.txt present (5)
• Renderability & JS (20) — share of visible text already in the server HTML; points lost above 50% and 75% JavaScript-only text
• Structure & readability (30) — H1, H2/H3 subheadings, lists/tables/FAQs, an early summary, all in the server HTML
• Metadata & consistency (25) — title, meta description, OpenGraph, structured data, canonical tag (5 each), in the server HTML

Markup added only by JavaScript does not count. Bands: 70+ Strong, 40–69 Needs work, below 40 At risk.

Ghost Test (hallucination risk)

Each question is asked twice: once using only the server HTML, once using the text visitors see after JavaScript runs. The model must answer only from the content or say "not found".

• "What is the pricing?" — asked on the pricing page (/pricing, /plans); skipped if none exists
• "What are the main features?" — asked on the features or product page (/features, /product)
• If no features or product page exists, one broad question is asked about the homepage instead

A question is flagged when the server-HTML answer is "not found" but the visitor-view answer is real. Hallucination risk is reported when at least half of the asked questions are flagged. A flag is evidence of risk, not proof that a specific AI product answered a real user incorrectly.

Full page: /methodology