Restaurant Growth Grader

POST /api/grader orchestrates:

• Business data — Google Places v1: details, opening hours, categories.
• Website — PageSpeed Insights (Lighthouse) plus a direct fetch of the restaurant's HTML.
• Benchmark — Google searchNearby for comparable venues in the area.
• Reputation — the Places reputation summary (rating, review count).
• Social — ScrapeCreators, only when a profile is confidently detected. Skipped otherwise, and the report says so.
• Narrative — one LLM completion that turns the numbers into prose.

Every step reports its own confidence and is recorded in the scan evidence, so a partial scan is still a usable report rather than a silent hole.

The website scanner fetches URLs a stranger typed, so lib/safe-fetch.ts sits in front of it: http(s) only, private / loopback / link-local addresses rejected, response body size capped, hard timeout. This is the one place in the estate where SSRF is a live risk, and it is guarded deliberately.

lib/providers/business-provider.ts counts Google Places calls against a daily and a monthly ceiling. Both ceilings come from env vars. Neither is set, so recordGooglePlacesUsage returns before it counts anything and the guard is a no-op.

Even once set, the counters live in process memory: they reset on every cold start and do not hold across serverless instances. Treat them as a speed bump, not a budget.

middleware.ts limits per IP with an in-memory counter, with the same caveat — one instance, one counter. The file itself flags the fix: an Upstash-backed limiter before this runs multi-region. The status app already has an Upstash connection it could share.

GRADER_AI_PROVIDER selects the provider at request time: openrouter (default), google, or anthropic. Each needs its own key. Selecting anthropic today throws at model init because ANTHROPIC_API_KEY is not set anywhere — the narrative then returns null and the report renders without prose instead of failing loudly.