A chatbot can display a real link beside an unsupported answer. Citation testing asks two separate questions: did the assistant use an authorized source, and does that source justify the particular statement? A link count alone cannot answer either question.
Related service: Custom AI chatbot development- 01Identify the factual claim
- 02Inspect its cited passage
- 03Check version and access
- 04Approve, qualify or reject the answer
SECTION 01
Specify what a useful reference includes
Agree whether citations point to a document, page, section or record. A long handbook link may be technically valid but unhelpful when the reader cannot locate the relevant paragraph. Preserve the source identifier and version even if the visible title later changes. Explain when an answer combines several sources rather than implying that one passage says everything.
Microsoft's RAG documentation describes grounding and source provenance. Ask for those links to be observable in your implementation, not merely present in a sales slide. Separate direct source facts from the assistant's summary or inference, especially when several conditions must be combined to answer the question.
SECTION 02
Illustrative example: conflicting return policies
Imagine a support assistant finds a current policy stating that clearance products are excluded from returns, alongside an old help article describing a general return window. Its answer says the customer can return a clearance item and cites the current policy. The citation exists, but it contradicts the answer.
The acceptance case should fail because the statement is unsupported, even though the link opens successfully. A better response explains the current exclusion, identifies the applicable source and offers support for exceptions. Add a second case where the customer has no access to an internal exception guide; the bot must not reveal that restricted content through its answer or citation label.
SECTION 03
Review evidence quality and access together
OWASP's authorization guidance calls for checks on each request. Apply that to opening a source as well as retrieving it. The citation path should not expose a private file through a public download link. Test revoked access, renamed files, removed pages and expired document links using an ordinary user account.
Ask reviewers to mark unsupported claims, missing conditions, stale versions and inaccessible links separately. Those categories lead to different repairs: source cleanup, retrieval changes, answer constraints or access controls. At handoff, keep a small set of annotated answers with the exact passages that support them. Recheck those cases after source or model changes rather than treating citations as a permanent guarantee.
SECTION 04
Frequently asked questions
Does a citation prove an answer is correct?
No. The passage must support the specific claim, be current enough for the task and be interpreted correctly. Check the evidence, not only the link.
Can the assistant cite internal documents?
Only when the user is authorized to receive the information and open the relevant reference. Enforce access independently of the model's wording.
Should every sentence have a citation?
Agree a useful rule for factual claims and combined conclusions. Repeating links after conversational filler can reduce readability without improving evidence.
PRIMARY REFERENCES
Sources and further reading
These references cover the standards, platforms, or published prices discussed in the guide. Worked examples and checklists are our editorial guidance.
- Retrieval-augmented generationMicrosoft Learn
- Authorization Cheat SheetOWASP
EDITORIAL METHOD
About this guide
We use AI to assist with drafting and editing. Catapult AI Work is responsible for the published content. Examples illustrate possible approaches; they are not client case studies unless identified as such.
Budget examples are not Catapult package prices. Check linked provider pages for current fees and plan limits before making a purchase.
Read the editorial policy