Documentation

Quick Start

Get started with CiteGuardian in three steps:

1
Sign Up

Create a free account at /register. You get 5 free credits to start.

2
Paste & Submit

Paste an AI-generated answer and add the source URLs or text it claims to cite.

3
Read the Verdict

CiteGuardian checks each claim against the sources and tells you what's supported, unsupported, or contradicted.

New Verification

From the Verify tab in the app, fill out the verification form:

  • Answer text — Paste the AI-generated text you want to verify. This is the content that references or cites sources.
  • Sources — Add one or more sources the answer claims to cite. Each source can be:
    • URL — CiteGuardian fetches the page content automatically.
    • Text — Paste the source text directly if it's behind a paywall or not publicly accessible.

Each source gets an ID (S1, S2, etc.) that maps to citation references like [S1] in the answer. Click Verify to submit. Each verification costs 1 credit.

Reading a Report

After verification completes, you'll see a detailed report with these sections:

Overall Summary
  • Supported rate — Percentage of claims that the sources actually support (e.g., "75% supported").
  • Total claims — Number of atomic factual claims extracted from the answer.
  • Decorative citations — Citations that don't meaningfully affect the verdict (see glossary).
  • Contradictions — Claims where the source evidence actively disproves the claim.
Claim Cards

Each claim is shown as a card with:

  • Verdict badgeSupported, Unsupported, or Unclear.
  • Confidence — How confident the verifier is in the verdict (0.0 to 1.0).
  • Citations — Which source IDs this claim references (e.g., [S1], [S2]).
  • Scrub test — Shows whether the cited evidence actually changed the verifier's judgment. A "decorative" flag means the citation didn't matter.
  • Notes — Explanation from the verifier about why it reached the verdict.
  • Contradiction flag — If the evidence directly disproves the claim, this is highlighted with a note explaining the contradiction.
Evidence Panel

Click a claim card to expand it and see the matched evidence chunks side-by-side with the claim. Each chunk shows which source and paragraph it came from (e.g., S1:C0).

Export

Reports can be exported as PDF or Markdown from the report page.

History

The History tab shows all your past verifications, ordered by most recent. Each entry shows a preview of the answer text, the supported rate, and the date. Click any entry to view the full report.

Batch Mode

For verifying multiple items at once, use the Batch tab. Upload a JSON file with an array of verification requests:

[
  {
    "answer": "The Earth orbits the Sun in 365.25 days [S1].",
    "sources": [
      {"id": "S1", "url": "https://example.com/astronomy"}
    ]
  },
  {
    "answer": "Water boils at 100C at sea level [S1].",
    "sources": [
      {"id": "S1", "text": "Water reaches its boiling point at 100 degrees Celsius under standard atmospheric pressure."}
    ]
  }
]

Each item in the array follows the same format as a single verification. Batch processing runs in the background, and you can track progress on the batch status page. Maximum batch size is 50 items.

Dashboard

Your dashboard is your home base after logging in. It provides:

  • Chrome extension — Download the extension to verify citations directly from any webpage with a right-click.
  • Credit balance — See how many credits you have remaining.
  • Upgrade plan — Subscribe to Starter ($19/mo) or Pro ($49/mo) for monthly credits and advanced features.
  • API keys — Create and manage API keys for programmatic access. Keys use the cg_ prefix.
  • Transaction history — Review all credit transactions (purchases, usage, refunds).

Glossary

Supported
The source evidence confirms the claim. The verifier found matching information in the cited material.
Unsupported
The source evidence does not confirm the claim. The cited material doesn't contain information that backs up what's being said.
Unclear
The verifier can't confidently determine whether the source supports or contradicts the claim. The evidence may be tangentially related or ambiguous.
Decorative Citation
A citation that looks legitimate but doesn't actually influence the verdict. The scrub test detects these by running the verifier twice — once with the evidence and once without. If removing the evidence doesn't change the verifier's confidence (delta < 0.10), the citation is flagged as decorative. This is the core of what CiteGuardian calls "citation laundering."
Contradiction
The source evidence actively disproves the claim. This is worse than unsupported — the cited material says the opposite of what the claim states.
Confidence Score
A value from 0.0 to 1.0 indicating how certain the verifier is about its verdict. Higher values mean stronger certainty.
Evidence Scrubbing (Scrub Test)
CiteGuardian's technique for detecting decorative citations. It judges each claim twice: once with the cited evidence (p1) and once without it (p0). The delta (p1 - p0) measures how much the evidence actually mattered. A small delta means the citation was just window dressing.
Pass/Fail Gate
An API mode that returns a simple pass or fail verdict instead of a full report. Used by RAG pipelines to block hallucinated responses before serving them. A response passes when the supported rate meets the threshold (default 70%) and no contradictions are found.