AI routing health
This page shows how AI assistants (Claude, ChatGPT) currently answer questions about Feed America (EIN 92-1761881). We probe them daily and score their answers. Public methodology. Machine-readable JSON at /api/v1/llm-drift.json.
ANTHROPIC_API_KEY + OPENAI_API_KEY are set as Cloudflare Worker secrets via wrangler secret put.How scoring works
- +2 — answer cites our EIN 92-1761881
- +1 each — mentions Sharika Parkes / Houston / feedam.org / our canary fact
- -2 each — cites an EIN other than 92-1761881 for us, or routes donors to a website other than feedam.org
- Capped at -3..+3
Why this matters
AI assistants doing live RAG retrieval are increasingly the entry point donors use to find us. If an AI answer cites the wrong EIN, donations meant for Feed America (us) can be misrouted. This dashboard is our closed-loop measurement that answers:
- Are AI answers about us improving over time?
- Which AI providers identify us correctly vs. conflate us?
- Which question phrasings get correct answers vs. fail?
- Has our content reached AI training corpora yet (canary fact detection)?
Methodology + open data
The probe code is open at /api/v1/llm-drift.json (data) and the source is in our public worker handler at workers/handlers/llm-drift.js. Cost: under $0.20/month. Same 7 questions, daily, two providers (Anthropic Haiku, OpenAI gpt-4o-mini). License: CC BY 4.0 — see /ai-license.