Citations are what separate a trustworthy AI assistant from a black box. When your chatbot shows users exactly where its answers come from, they can verify the information and trust the response. This guide explains how citations work and how to optimize them for accuracy.
1. What Are Citations?
Citations are references to the source documents that informed the AI's response. When enabled, they appear as footnotes below each answer, showing users exactly which documents were used.
Example Citation Display
Our return policy allows full refunds within 30 days of purchase. After 30 days, we offer store credit for items in original condition.
Footnotes:
Return-Policy-2024.pdf(score: 0.94) https://company.com/returns(score: 0.87)
Why Citations Matter
Build trust
Users can verify answers by clicking through to source documents. This transparency builds confidence in your AI assistant.
Catch errors
If the AI cites the wrong document, you know there's a problem with your knowledge base or the question was ambiguous.
Guide improvements
Relevance scores show which documents are being retrieved. Low scores indicate your knowledge base may need better content.
2. How Citations Work
When a user asks a question, Chat.co uses a sophisticated retrieval process to find relevant information and attribute sources.
The Retrieval Process
Hybrid Search
Your question is analyzed using both semantic search (understanding meaning) and keyword search (matching exact terms). This hybrid approach finds relevant content even when users phrase questions differently than your documents.
Document Retrieval
The system retrieves 15-30 potentially relevant document chunks from your knowledge base based on the search results.
Reranking
Retrieved documents are reranked using advanced AI models to identify the 5-10 most relevant pieces. This step dramatically improves citation accuracy.
Response Generation
The AI generates an answer based on the top-ranked documents, tracking which sources informed each part of the response.
Citation Display
Sources are displayed as clickable footnotes with relevance scores. Users can click to view or download the original documents.
What Gets Cited
- Uploaded documents — PDFs, Word docs, text files you've added
- Crawled web pages — URLs you've imported into the knowledge base
- Q&A pairs — Custom question-answer training data
3. Citation Settings
You can control whether citations are displayed to users.
Enabling/Disabling Citations
- Go to your chatbot's dashboard
- Click
Appearancein the sidebar - Find the Citations toggle
- Enable or disable as needed
- Save changes
Citations
When enabled, answers include footnotes and source references (where available).
When to Disable Citations
While citations are generally recommended, you might disable them for:
- Simple FAQ bots — Where the source is obvious
- Cleaner UI — For consumer-facing chatbots where footnotes feel too formal
- Confidential sources — When you don't want to reveal document names
Recommendation: Keep citations enabled for support, knowledge base, and internal assistant use cases. The transparency helps users trust the information.
4. Improving Citation Quality
The quality of your citations depends on the quality of your knowledge base. Here's how to optimize for better source attribution.
Document Structure
Use clear headings
Structure documents with H1, H2, H3 headers. This helps the AI understand content hierarchy and retrieve the most relevant sections.
Keep sections focused
Each section should cover one topic. Long, rambling sections make it harder for the AI to pinpoint the relevant portion.
Include key terms
Use the terminology your users will search for. If users say "cancel" but your document says "terminate," add both terms.
File Naming
Document names appear in citations. Good names help users understand the source at a glance.
- doc1.pdf
- final_final_v3.docx
- untitled.txt
- Copy of policies.pdf
- Return-Policy-2024.pdf
- Employee-Benefits-Guide.docx
- API-Authentication-Guide.txt
- Pricing-Enterprise-Plans.pdf
Content Quality
| Do | Don't |
|---|---|
| Write complete, self-contained answers | Reference other documents ("see policy X") |
| Use text-based PDFs with searchable text | Upload scanned images without OCR |
| Remove duplicate documents | Keep multiple versions of the same content |
| Update outdated content regularly | Leave obsolete information in the KB |
| Break large documents into logical sections | Upload 500-page manuals as single files |
Knowledge Base Hygiene
- Remove duplicates — Multiple documents with similar content confuse citation attribution
- Delete outdated content — Old policies or deprecated information can lead to wrong citations
- Consolidate related content — Better to have one comprehensive document than five fragments
- Test regularly — Ask questions and verify citations point to the right sources
5. Understanding Relevance Scores
Each citation includes a relevance score from 0 to 1, indicating how closely the source matched the query.
Score Interpretation
Strong semantic match. The document directly addresses the question.
Relevant content that partially answers the question or provides context.
Somewhat related content. May indicate the knowledge base lacks better sources.
Loosely related or fallback content. Consider adding better source material.
Using Scores for Improvement
Monitor citation scores to identify knowledge gaps:
- Consistently low scores — Add more content on that topic
- Wrong documents cited — Improve document titles or restructure content
- No citations — The topic isn't covered in your knowledge base
6. Troubleshooting Poor Citations
Common citation issues and how to fix them.
No citations appearing
- Check that citations are enabled in Appearance settings
- Verify your knowledge base has content (Sources page)
- Ensure documents have been processed successfully
- Test with a question you know is covered in your documents
Wrong documents being cited
- Remove duplicate or near-duplicate documents
- Use more specific document titles
- Restructure content so each document has a clear focus
- Add the missing correct content if it doesn't exist
Low relevance scores
- Add more content on the topic being asked about
- Use terminology that matches how users ask questions
- Break up large documents into focused sections
- Ensure PDFs have searchable text (not just images)
Citations link doesn't work
- Document download links expire after 15 minutes—refresh the chat
- Check if the source document was deleted from the knowledge base
- For URLs, verify the original webpage is still accessible
Pro Tip
Create a test document with known content and test questions. This helps you verify citations are working correctly and gives you a baseline for comparison when troubleshooting other documents.