Methodology ยท v1.0
Exactly how the GEO Score is computed: how queries are generated, how we detect mentions in AI answers, the scoring formula, and how fixes are ranked. Transparent and auditable.
Versioned. Refreshed quarterly. No black-box ranking.
01 ยท Query generation
Every scan generates exactly 8 queries split across four categories: 2 general (โbest X providersโ), 2 local (โbest X in Yโ), 2 use-case (industry-specific), and 2 competitor-alternative (โalternatives to Xโ). Missing inputs (no region, no competitors) are backfilled with extra general or use-case queries to keep the budget at 8.
Query templates
What are the best $${industry} companies?Top $${industry} providers in 2026Best $${industry} in $${region}$${industry} near $${region}industry-specific template (per catalog)industry-specific template (per catalog)Alternatives to $${competitor}$${industry} similar to $${competitor}02 ยท Mention detection
Each engine response is scored 0โ100 along five dimensions: exact name match, URL host match in citations, fuzzy name match for typos and tokenization variants, description proximity bonus, and a small position bonus. We use a regex first-pass, then a Levenshtein-based fuzzy match, then an entity disambiguation step that resolves common name collisions.
Per-response score components
The per-engine score is the mean across the 8 queries, clamped to 0โ100.
03 ยท Scoring formula
GEO_Score =
ย ย 0.40 ร ChatGPT
+ 0.25 ร Gemini
+ 0.20 ร Perplexity
+ 0.15 ร AI Overviews
Bands are deterministic:
70โ100
40โ69
15โ39
0โ14
04 ยท Fix prioritization
Each of the 30+ fixes carries an effort score (1โ5) and an estimated impact score (1โ5). The base priority is computed as priority = (impact ร 1.0) โ (effort ร 0.4). Diagnostic flags from site signals (no schema, no llms.txt, no Google Business Profile) override the ranking for missing-essentials cases โ those fixes float to the top regardless of base priority. The top 5โ8 fixes are returned.
05 ยท Engine-specific notes
Heavy weight on schema and review aggregators. Reddit and StackExchange answers contribute disproportionately. New entrants enter the answer set within ~2 weeks of consistent mentions.
Pulls heavily from Google Knowledge Graph, AI Overviews signals, and Google Business Profile. Local queries lean on GBP completeness more than other engines.
Citation-first. Inbound links from reputable domains (DA 70+) and explicit press placements move the needle fastest. Schema is necessary but not sufficient.
Generative answers above SERP. Performance closely tracks classic SEO + structured data. Mobile-first crawl and Core Web Vitals matter materially.
06 ยท Stub-vs-live disclosure
When live engine API keys are configured, scans hit the real OpenAI, Gemini, Perplexity, and AI-Overviews endpoints. When they are not, we run a deterministic synthetic engine that mimics realistic behavior using the catalog and your business signals โ useful for demos and previews, not a substitute for a live scan. Live versus stub mode is reported in your scan record so you can audit it later.
07 ยท Versioning + sources
Engine weights and the fix library are refreshed every quarter. Reference data: SimilarWeb and SparkToro AI-search-traffic share, public engine documentation, plus internal buzzi.ai client analytics. Engine outputs are stochastic; same scan re-run can vary slightly. Re-scan in 30 days to gauge fix impact.