Methodology

Last updated: July 2026

Seencite is built on one principle: an AI-visibility number is only worth showing if it's honest about its own uncertainty. This page documents exactly how we measure, score, predict, and prove — so you can judge the numbers, not just trust them.

What we measure — three pillars

• SEO — classic crawlability and on-page signals.
• AEO (Answer Engine Optimization) — whether your pages are structured so an AI can extract and cite them.
• GEO (Generative Engine Optimization) — whether AI answer engines actually mention and cite you when someone asks.

The GEO score (0–100)

The GEO readiness score is a weighted, severity-adjusted pass-rate across five subscores. The weights are fixed and public:

• Crawler access — 30%: can AI retrieval bots (GPTBot, PerplexityBot, Google-Extended, etc.) actually fetch your pages?
• Extractability — 25%: is the substance answer-first and readable without running JavaScript?
• Structured data — 20%: Organization, Product, FAQ and Article schema an engine can quote with confidence.
• Renderability — 15%: does key content appear server-side, and does the page render fast enough to be indexed?
• Entity clarity — 10%: is your brand connected to the knowledge graph (Wikidata, dense sameAs)?

The score is a prioritization tool, not a prediction of traffic. It tells you where the points are, ranked by severity.

How we query the AI engines

We ask each answer engine real category questions and read whether your brand is mentioned and whether it's cited (linked). Engines are queried the way they actually answer, and every result is labeled by retrieval mode — live web, grounded search, or training data — because a mention in one is not the same as a mention in another. An engine whose calls all error is reported as unavailable, never as 0%.

Honest metrics — the part most tools hide

AI answers vary every time you ask. Reporting a single confident-looking rank hides that variance. Instead, for every rate we run the prompt set multiple times and report:

• A point rate (e.g. "51% mentioned"),
• A 95% confidence interval computed with the Wilson score interval (well-behaved for small samples and rates near 0 or 100%),
• The sample size (n) the interval is based on.

So a real number reads: 51% mentioned · 95% CI 44–58% · n=96. When two intervals overlap, treat the difference as noise, not signal. That's the whole point.

Predict, then prove

The what-if simulator models the lift you might get from changes you're weighing (more entity links, schema fixes, answer-first pages). Its output is explicitly labeled a model estimate — never a promise. After you ship fixes, we re-measure and show the realized before→after change. The estimate is a guide; the proof is measured.

What we deliberately do not do

• We don't invent a metric we can't measure. If we track only the sources where you're absent, we won't dress that up as a "cited on X of Y" ratio.
• We don't present a projection as a guarantee.
• We don't show an attribution number we haven't wired yet — it stays "coming soon" until it's real.
• We don't score question-shaped headings as substance; extractable evidence beats formatting tricks.

Limitations

AI engines change frequently and answer non-deterministically. Small samples carry wide intervals — that's why we show them. Numbers are directional evidence for decisions, not precise forecasts. When in doubt, re-run and watch the interval, not the point.

Questions about the method? support@seencite.com.