Summary

A new Site Audit feature (under Content Optimization) that scores any page against a 47-signal AEO/GEO rubric and returns page-specific, AI-written fix recommendations with ready-to-paste drafts.

The rubric is an open standard implemented entirely in-house — no third-party scoring API.

How it works

POST /api/audits {url, brandId} returns a running row immediately; the audit runs in the background and the client polls GET /api/audits/:id until it flips to completed/failed. Per audit:

1. Fetch the page via Scrape.do (proxy + JS render); robots.txt / llms.txt fetched directly with a Scrape.do fallback only when blocked.
2. 33 deterministic signals — cheerio DOM parsing (schema/headings/lists/tables/meta/links, Flesch readability, etc.).
3. 13 LLM-judged signals — one batched structured-output call. The page is evaluated on its own inferred topic — no brand/external context is injected, so auditing an arbitrary URL never imports unrelated subject matter.
4. Wikidata brand-entity lookup.
5. AI fix recommendations — a second LLM call over the fixable failing signals, returning prioritized advice + drafted snippets (meta description, H1, BLUF, FAQ JSON-LD, simplified rewrites…).

Score = weighted category mean, normalized over evaluated signals (coverage shown as X/47).

Guardrails

• Feature-gated behind content_optimization (server-side, not just UI).
• Monthly per-org quota — Starter 100 / Growth 500; Enterprise & self-hosted unlimited. Enforced at request time; usage recorded only for completed audits (failures don't burn quota). Mirrors the existing brief-usage pattern.
• LLM model is provider-flexible via resolveModel — env AUDIT_LLM_MODEL (default google/gemini-3-flash-preview); self-hosters can point it at any provider with a key.

Web UI

Site Audit page: circular score gauge + grade, pass/warn/fail summary chips, category breakdown bars, signals grouped by category (each expands to finding + why + AI-written fix with copy), an AI Recommendations card, recent-audits list with delete, an issues-only filter, staged loading state, and an empty state.

Schema

• 00017_site_audit.sql — site_audits + audit_signal_results (+ RLS, org-membership scoped).
• 00018_site_audit_recommendations.sql — recommendations jsonb column.
• 00019_site_audit_usage.sql — site_audit_usage quota table (RLS, service-role only).

New env vars (see server/.env.example)

• SCRAPEDO_API_KEY — page fetching.
• AUDIT_LLM_MODEL — defaults to google/gemini-3-flash-preview.

Notes

• Adds cheerio to the server package.
• Server-only runtime — needs a server redeploy after merge. The three migrations are already applied to production; self-hosters get them via the repo.
• Cost: ≈10k tokens/audit (~$0.016 at $0.5/$3 per 1M in/out).