AI Sourcing & Verification Lab

Can CFI.co guarantee that an AI system will cite your organisation? No. AI answers vary by system, date, location, prompt and user context — no responsible publisher can guarantee future selection or citation. What we can do is show you the evidence: the exact questions asked, the systems and models, the dates, the outcomes (including the negative ones), and the tools to repeat every test yourself.

This page is the index to the CFI.co AI Evidence Library — a dated, append-only record of how AI systems select and cite CFI.co content. Every case below links to its raw evidence: the exact prompt sent, the system and model, the full unedited response, every cited URL, the HTTP status and the stated limitations. The library is covered by the same GPG-signed manifest and append-only git history as the article records.

Machine-readable index: /archive-data/ai-evidence.json · Raw case files: ai-evidence/

Published cases (2)

A failure, documented: Perplexity API quota exhaustion and recovery (24 July 2026)
Systems tested: Perplexity (sonar).
This is a transparency/negative-result case, not a citation claim.
Stated limitations: The failure was on CFI.co's side (test-account billing), not Perplexity's service; Error bodies are token-redacted as a credential-hygiene precaution.
Evidence record — 1 data file, manifest and limitations.

Live citation test: 8 questions x 4 AI engines (24 July 2026)
Systems tested: ChatGPT (gpt-4o-search-preview), Gemini (gemini-2.5-flash), Perplexity (sonar), Claude (claude-haiku-4-5-20251001).
Selected (level 3): engines returned CFI.co pages in citation metadata. Use of content in answer text (level 4) is visible in results.json answers but not systematically scored here.
Stated limitations: Single dated run; results change over time; API access mode, not the consumer applications; Citation detection = CFI-owned source in engine citation metadata (see limitations.md); No claim of endorsement, evaluation or ranking.
Evidence record — 1 data file, manifest and limitations.

In preparation (2)

2 further cases are being assembled as full evidence packs. Until a pack is published it is a placeholder that makes no evidentiary claim, it is not counted among the published cases above, and the organisation concerned is not named — naming a client for a case we cannot yet show would be the claim we are trying to avoid making. The placeholders are visible in the public repository with status: pending_documents.

What we do not claim

We do not claim guaranteed citation, ranking or treatment by any model; revenue or transaction attribution; equivalence between sponsored content and independent earned journalism; or that cryptographic hashes prove the truth of underlying claims — they prove the record has not been altered. “Cited” means an engine returned a CFI.co-owned source in its citation metadata. It never means “endorsed”, “evaluated” or “ranked”.

Evidence maturity

Each case states the maturity level it demonstrates and no more: 1 Accessible → 2 Discoverable → 3 Selected → 4 Used → 5 Accuracy contribution → 6 Commercial contribution (only ever claimed with client-supplied data).

How we test

Queries are sent server-side to each engine’s public API from CFI.co’s own infrastructure. API results are not identical to the consumer apps — no login state, no location personalisation, no interface context — so treat these as one measurement mode and your own in-app tests as another. A control question with a known, stable citation validates the detector on every automated run, and a negative control confirms the detector does not report citations that were not made. Failures are published, not curated away.

The verification chain

Every case connects to a chain anyone can walk: published article → visible and machine-readable sponsorship classification → hashed JSON record on GitHub → change history → AI test → archived result.