How the Found Score is calculated.

Fetches the homepage like a crawler. Checks status, content type, and server-rendered text. Note: Google renders JavaScript (with delay); most AI crawlers do not — thin raw HTML is scored as a warning for exactly that reason, not as a Google problem.

Parses robots.txt with longest-match rules. Distinguishes answer-engine bots (OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot, Bingbot — blocking these removes you from live AI answers, penalised) from training bots (GPTBot, ClaudeBot, CCBot — blocking these is a legitimate licensing choice, noted but never penalised).

Fetches the declared or conventional sitemap, validates the urlset, reads the newest lastmod (treated as a hint, as engines do). Warns past 183 days.

Checks meta robots and X-Robots-Tag headers for accidental noindex.

Probes www and http variants; expects single-hop redirects to one canonical origin. Unresolvable variants are acceptable.

Parses JSON-LD for Organization/LocalBusiness + WebSite with name, url, logo. Framed as entity clarity (site names, knowledge panels, AI identity-matching) — not as a ranking factor, because it isn’t one.

Detects the brand from schema, checks homepage and sampled page titles carry it.

Discovers the about page via sitemap + link inventory (soft-404s rejected), measures visible text. Engines resolve “who is this business” largely here.

Planned — Places API.

Planned — site vs Places listing.

Planned: monthly prompt runs against ChatGPT and Perplexity APIs using customer-question archetypes; records whether the brand is named or cited, and by whom instead.

LLM-judged (Claude, versioned prompt, cached by content hash for consistency): counts sourced statistics, named quotations, citations; flags keyword stuffing. The point mapping from counts is deterministic code, not model whim. Basis: Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024 — sourced statistics, quotations and citations raised generative-engine visibility 30–40%.

Word counts across sampled pages plus 5-word-shingle similarity to catch near-duplicate template pages. Framed as substance and duplication risk — not “word count is a ranking factor”, which it isn’t.

LLM-judged: definitions, procedures, comparisons, lists/tables, modular headings — five subscores of 2. Explicitly not “has an FAQ block”. Google states no special formatting is needed for its AI features; this check targets the other engines, which demonstrably lift structured passages.

One clear h1 (multiple h1s are valid HTML5 and fine for Google — docked one point only for outline ambiguity, never called an error), no skipped levels, descriptive text.

Planned — referring-domain summary.

Planned — Reddit/UGC/directory mentions. Brands are cited via third-party sources far more than via their own domains.