Method & ruleset
AEO/GEO signals shift faster than classic SEO. The check catalogue, engine list and access methods are versioned, dated config — and every report records the version it was produced with.
Ruleset 2026.08.2Method 2026.08.3
What this product does not promise
We measure readiness and visibility. We never promise citations. AI answers are not a ranked list, so there is no rank to report; answers vary by run, locale and model version, which is why every visibility figure is a rate over repeated samples with a margin of error. Published GEO uplift percentages come from a thin research base — Aggarwal et al., “GEO: Generative Engine Optimization”, arXiv:2311.09735 (2024) — and are labelled directional wherever we show them.
Audit dimensions and weights
Ruleset 2026.08.2, dated 2026-08-19. D1 (access) and D3 (extractability) carry the most weight.
| Dimension | Gate | Tier | Weight |
|---|---|---|---|
| D1 AI crawler access & retrievability Can AI engines fetch this site at all? | retrievability | 1 | 30 |
| D2 Machine-readability & structured data Can an engine parse what this page is about? | extractability | 2 | 15 |
| D3 Content structure & extractability Can an engine lift a clean answer out of the page? | extractability | 2 | 25 |
| D4 Authority & E-E-A-T Is this a source an engine should trust? | trust | 2 | 15 |
| D5 Entity clarity & GEO Does the brand resolve to a clear, verifiable entity? | trust | 3 | 8 |
| D6 Freshness Is the content current enough to be reused? | trust | 2 | 7 |
How the score works
Deterministic: the same checks always produce the same score.
- Each check scores Pass = 1, Warn = 0.5, Fail = 0. A check that could not run is excluded — never counted as a pass.
- Dimension scores are averaged, then combined by the weights above.
- Tier-1 gate: any failed Tier-1 blocker caps the overall score at 59. A site cannot be A-grade AI-ready while it blocks GPTBot.
- Site-level results aggregate pages by weighted mean, but take the worst case for blockers.
- Readiness and visibility are never blended into one number. One is a diagnostic score, the other a measured outcome.
Measured vs heuristic
Every finding says which it is.
Measured means directly observed — robots.txt says X, the edge returned a 403, the engine cited this domain. Heuristic means inferred by a rule of thumb — that a page “reads answer-first”, or that a quotation is attributed. An engine answer’s sentiment (recommended, dismissed or mentioned) is the same kind of heuristic — a lexical read of the wording around the mention, not a measured fact. Heuristics are useful signals, not facts, and are labelled so nobody has to guess which they are looking at.
AI user agents we report on
Blocking one of these is the single most common invisible cause of absence. What each one governs matters — advising an unblock of the wrong bot is worse than useless.
| Token | Vendor | Blocker | What it governs |
|---|---|---|---|
| GPTBot | OpenAI | Tier 1 | Crawling for ChatGPT’s index and model training. |
| OAI-SearchBot | OpenAI | Tier 1 | Surfacing your pages as sources in ChatGPT Search results. |
| ChatGPT-User | OpenAI | Advisory | Live fetches when a user asks ChatGPT to open your page. |
| PerplexityBot | Perplexity | Tier 1 | Indexing for Perplexity answers and citations. |
| Perplexity-User | Perplexity | Advisory | Live fetches triggered by a Perplexity user. |
| Google-Extended | Tier 1 | Gemini and Vertex AI grounding/training only. It does NOT control AI Overviews indexing — that follows Googlebot, so blocking Google-Extended will not remove you from AI Overviews, and unblocking it will not add you. | |
| ClaudeBot | Anthropic | Tier 1 | Crawling for Claude’s index and training. |
| anthropic-ai | Anthropic | Advisory | Legacy Anthropic crawler token; still honoured by many sites. |
| Bytespider | ByteDance | Advisory | ByteDance/Doubao crawling. |
| Amazonbot | Amazon | Advisory | Alexa and Amazon AI answer surfaces. |
| Applebot-Extended | Apple | Advisory | Apple Intelligence training. Applebot itself still handles Siri/Spotlight indexing. |
| Meta-ExternalAgent | Meta | Advisory | Meta AI crawling. |
| CCBot | Common Crawl | Advisory | Common Crawl — a training corpus many engines derive from. |
Engine access methods
Visibility method config 2026.08.3, dated 2026-08-28. Where the official API does not reflect what a real user sees, UI capture is preferred — and the method used is recorded on every data point.
| Engine | Tier | Preferred method | Citations | Notes |
|---|---|---|---|---|
| Perplexity | primary | API | yes | Closest to purpose-built: the API natively returns the sources behind the answer. |
| ChatGPT | primary | UI capture | yes | API output does not match the consumer app: search mode, memory and personalisation differ. UI capture is preferred for fidelity; the API is the cheaper fallback. |
| Gemini | primary | API | partial | The model API is a distinct surface from Google Search; citations are inconsistent. |
| Claude | primary | API | yes | Web-search tool use returns the sources consulted. |
| Google AI Overviews | primary | UI capture | yes | No API exists. Scrape-only and best-effort: overviews are auto-triggered by query type, vary by locale, and the surface is fragile and ToS-sensitive. |
| Microsoft Copilot | extended | UI capture | yes | Higher tiers only. |
| Grok | extended | API | partial | Higher tiers only. |
Crawl politeness and safety
What the crawler will and will not do.
- Only publicly accessible pages are fetched. Authentication and paywalls are never bypassed.
- The crawler identifies itself, rate-limits per host and caps pages per host.
- The fetcher refuses internal, loopback, link-local and cloud-metadata addresses, and any scheme that is not http or https — validated at the socket, so a DNS rebind or a redirect cannot slip past it.
- Response size and redirect depth are capped, and fetched content is rendered as text, never as markup.