TL;DR. Every blueprint is generated on-demand by grounding a frontier LLM (Gemini 3 Flash via Lovable AI Gateway) in first-party country profile data — currency, payment rails, supplier hubs, marketplace share, regulations — then rewritten into extractable, question-first sections and cached with a 30-day freshness cycle.
Every generation call combines three inputs: (a) the founder's brief (country, budget, vertical, skills), (b) a country profile record from our country_profiles table, and (c) a structured prompt template that enforces question-first H2s, 40–60 word direct answers, and a Sources & methodology footer on every blueprint.
Country profile fields (per market): currency + symbol, dominant payment rails (e.g. bKash, UPI, Pix), common supplier hubs, top ecommerce marketplaces, business registration path, tax regime, common regulatory gotchas, and typical unit economics anchors. These fields are the ground truth the model is forced to cite instead of inventing generic advice.
We use google/gemini-3-flash-preview via the Lovable AI Gateway for blueprint generation and workspace chat. Every prompt includes the country profile as MARKET context and forbids "consult a professional" / disclaimer language — the model must answer as an operating partner.
Outputs are streamed and structured (TL;DR block → question H2s → bulleted supporting evidence → Sources footer) so answer engines can extract self-contained blocks.
Programmatic SEO pages (/business-ideas/<slug>) are cached in seo_pages after first generation. Any request older than 30 days triggers a background regeneration on next hit, so the corpus never grows stale relative to local market shifts.
Shared blueprints (/b/<slug>) display a visible Last updated date and can be regenerated by the owner from the dashboard.