What we measure (and what we don't promise)

High confidence — direct measurement

These checks are deterministic: we read what's there and report it. If the audit says you're blocking OAI-SearchBot in robots.txt, you are. If the bot access test gets a 403, the bot can't reach you.

• ✓ robots.txt parsing — exact rules per user-agent.
• ✓ AI bot access (CDN/WAF) — actual HTTP status returned to each bot's user-agent.
• ✓ SPA / SSR ratio — bytes of content visible without JavaScript vs. with.
• ✓ Schema.org / JSON-LD presence — what's in the HTML right now.
• ✓ Technical basics — TTFB, HTTPS, sitemap, canonical, redirect chain.
• ✓ Citation snapshots — when AI search returns sources, we report those URLs verbatim.

Medium confidence — observed correlation

These are patterns we see, not laws we've proved. The data is real; the inference is reasonable but not validated.

• ◐ "Page-type playbook" in source maps — descriptive of which URL patterns get cited today, not a prescription for ranking. Top domains may be cited because they have authority, not because they use /blog/ URLs.
• ◐ Topic clustering (buying / how-to / comparison) — heuristic regex on query keywords. Works for English, breaks on ambiguous or non-English queries.
• ◐ Domain rollup ranking — accurate for the queries we ran. Rankings shift week to week.
• ◐ "Why am I (not) cited?" reasoning — Claude's plausible explanation post-hoc. It can't see the actual retrieval algorithm; treat it as a brainstorm.

Low confidence — heuristics from traditional SEO

We borrow these from established SEO practice and apply them to AI search. They might transfer cleanly. They might not. There's not enough public research yet either way.

• ○ llms.txt — spec is new (2024–2025). Anthropic respects it; OpenAI partially; Perplexity unclear. We surface its presence but don't claim it lifts citations.
• ○ E-E-A-T signals (author, date, Person/Org schema) — comes from Google's quality guidelines. Reasonable to assume LLMs prefer authoritative content; not empirically validated for citation rates.
• ○ "Recommended schemas" per page type — informed guesses based on what makes sense for AI parsing. Adding FAQPage may or may not move the needle.
• ○ The 0–100 score itself — our weighting (retrieval 30, SPA 25, llms.txt 15, schema 15, technical 10, E-E-A-T 5) is sensible but arbitrary. The score is useful as a proxy, not as an industry benchmark.

Generated content — input, not answer

Anything Claude or another LLM produces for you is an opinion based on what it observed. We never claim AI-generated output will itself be cited.

• ✎ Content briefs (title, outline, FAQs, recommended sources) — Claude's synthesis from the search step. Good editorial input, not a guaranteed-to-rank specification.
• ✎ First-draft markdown — Claude writing prose. Not publish-ready. Treat as a starting point for a writer to edit, fact-check, and add real examples.
• ✎ Competitive gap analysis — Claude reading a top URL and listing what it covers / doesn't cover. Useful for differentiation; not a ranking factor.
• ✎ Pre-generated JSON-LD — technically valid markup. Adding it doesn't guarantee anything beyond making your structured data parseable.

Where the real value sits

The strongest part of the product is the measurement loop: snapshot what AI cites today, change something on your site, snapshot again, see if it moved. That's defensible regardless of which heuristics turn out to be right. We're building toward better empirical grounding as the field matures.

We publish what we observe across all audited domains every month at /insights — confidence-tiered, with full caveats. No other GEO tool does this.

If you have data — public or private — that strengthens or contradicts any of the above, we want to hear it. Email hello@citeai.io.