Four separate responsibilities
- Facts: normalized business records with source and collection information.
- Rules: deterministic conditions applied to your answers, linked to official sources.
- Analytics: calculations over the filtered records.
- Explanation: evidence-based observations, optionally ordered by an AI provider.
Comparable business counts
Each matching record represents a business location, not necessarily a unique legal business. Closed and inactive records are excluded. Unknown status remains included. Filters use the normalized category, exact city, ZIP, website status, and optional minimum rating and review count. Categories are broad; a restaurant count is not a Haitian-restaurant count.
Ratings and reviews
Average rating is the arithmetic mean across locations with a recorded rating, without weighting by review count. Median reviews uses only locations with a recorded review count. Each metric shows its sample size. An empty sample produces no metric, never a zero rating.
Website adoption
Website percentage = locations with a recorded website ÷ locations with known website status. “No website” requires an explicit negative value in the source. Unknown locations are shown separately and excluded from this denominator.
Geography
Geographic distribution counts matching locations by recorded ZIP. Missing ZIPs stay in an explicit unknown group. Counts are not density: there is no land-area or population denominator in V1. Neighborhood boundaries, radius search, heatmaps, and opportunity scores require additional verified inputs.
Rules and confidence
“Required” means a source-backed condition matches the answers you supplied. “Likely required” means more details may be needed. “May apply” calls for verification. “Recommended” is a planning or technology suggestion. Savannah-specific requirements are enabled only after you confirm the operating location is within city limits. No property is approved by this platform.
Reproducible reports
Reports preserve the input answers, derived analytics, supporting sources, and a hash of the dataset content. Checklist completion is stored separately. Running research later can produce different results as the underlying dataset changes.
The AI boundary
V1 parses supported category, city, ZIP, and no-website questions into validated filters. An optional language-model provider receives supported aggregate observations and chooses their display order from an allowlist. It cannot produce new factual prose, generate SQL, or change legal requirements. If the provider is unavailable, the app returns the deterministic data summary.