A citation under an AI answer is a promise that can be tested. It should help a visitor verify a material claim and help an operator diagnose a bad answer. That makes citation testing a practical launch task, not a design detail.
Audit the citation and the answer together
Testing only whether a link opens misses the central risk. A live source can still be irrelevant, old, or too broad for the claim. Review each answer as a pair: what does it say, and what does the linked page actually establish? If a chatbot says “next-day delivery is available,” the source must cover the destination, cut-off time and product conditions the visitor would reasonably assume.
Create a small audit sheet with the question, expected source, answer, cited URL, verdict and notes. Assign a verdict such as supported, partly supported, unsupported or wrong source. This avoids the common mistake of passing an answer simply because it sounds sensible.
Build a test set that invites mistakes
Start with real questions from sales and support, then add questions designed to test boundaries. Use indirect wording, shortened questions and conditions that appear in the source. For a fictional bike shop, “Can I send back a helmet?” is not enough. Also ask “I removed the tags but never wore it—does the 30-day return still apply?” and “Can you extend the return period for a gift?” The latter may need a person even if the general policy is clear.
| Category | What to check |
|---|---|
| Direct factual question | The answer cites the expected product, policy or help page. |
| Qualified question | Important exclusions and conditions remain in the answer. |
| Ambiguous wording | The assistant asks for context or stays within what the source supports. |
| Absent information | It does not guess and offers an appropriate next step. |
| Changed content | The answer and citation reflect the current approved material after an update. |
Keep the expected answer brief. The target is not identical prose; it is a faithful response with evidence. A different but equally valid source may be acceptable when it supports the claim better than the one you initially selected.
Check failures where they happen
When a test fails, classify it before editing content. A no-match case may be a genuine knowledge gap. A result with an apparently relevant candidate but no useful answer may be a retrieval problem: the knowledge exists but the wording was not found. A cited but incorrect answer can point to stale or conflicting source material. Each diagnosis needs a different repair.
RobiFox’s missing-knowledge view explicitly distinguishes a zero-match gap from a question that found candidates but did not produce an answer; the latter can be linked to existing knowledge so the visitor’s wording becomes searchable. That is a useful operating distinction because creating a second answer for the same fact can create a future conflict. See the related workflow in How it works.
Validate citation links separately
Run a link check across the citations produced during testing. Confirm that each URL resolves for a normal visitor, points to the intended page and does not require a private login. Then inspect the visible title and surrounding text. A source can be technically reachable yet become unhelpful after a website redesign.
RobiFox’s public methodology makes live cited-source links a distinct check, alongside answerability and refusal tests. It also says that an internal result is not a prediction for a customer site. That separation is sensible: a citation system can be sound while a particular website still contains vague or contradictory content. Read the stated scope in How we measure.
Re-test after content changes
Put high-impact sources on a simple change list: prices, shipping rules, returns, availability, warranty and contractual terms. Whenever one changes, rerun the questions that depend on it and a few neighbouring questions. Do not assume a changed page affects only the obvious question; an updated delivery rule may also change “When will my order arrive?” and “Can I ship to an island?”
RobiFox says it re-reads a site weekly and sends genuinely changed pages back for processing, but an automatic refresh is not a substitute for your acceptance test. The business still decides whether a new policy is clear, approved and ready for visitors. The review queue is intended to surface conflicts and retain a human decision when sources disagree.
Use the conversation record as audit evidence
After launch, sample completed conversations rather than relying only on synthetic tests. Look for citations that visitors ignore, answers that cause a follow-up, repeated reformulations and negative ratings. The source record should let a reviewer see what was asked, what was answered and which source was referenced. That makes an incident actionable: fix the statement, fix the source, or record that the question needs human judgement.
Source citations are most useful when they create this feedback loop. They give a visitor a way to check a claim, and give the team a precise place to improve it. For the broader set of controls, see RobiFox features.
For the next step, measure overall chatbot accuracy.
Cover image: AI-generated editorial illustration, not a product screenshot.
Checking a real file citation in RobiFox
Our 1 October 2026 Hungarian help test asked which document formats can be uploaded. The assistant answered, translated here into English: “You can upload a PDF, Word or spreadsheet file as new material.” It cited RobiFox_Kezikonyv.docx, section 39, together with section 38. The manual’s document-upload chapter covers those formats.
The original Hungarian question and response are in the downloadable v2 result. The file:// section reference is an internal document reference, not a publicly reachable web URL. Across all 40 questions with an expected source, 35 contained the exact expected reference. That does not establish the factual correctness of every answer.
Read the dated methodology and original responses or download the 52-question result (JSON). This measures our own Hungarian documentation, not English answer quality or customer results.
