Current guide
Citation and Hallucination Analysis
Citation and hallucination analysis
Citation and hallucination analysis checks a recorded response against the sources it cites. It extracts the factual claims from the response, tests each one against its citation, and reports the result claim by claim, alongside a live status check of every cited link. This page covers both layers: the automatic link verification and the on-demand analysis, plus the full report they build.
Prerequisites:
- A recorded response: The analysis runs on captured assistant turns. See chat capture.
- Cloud Sync & Intelligence on: AI features require it, along with an internet connection. See sign in to the extension.
- AI credits: The analysis draws on your monthly credit pool. See Account and Billing.
Two layers, two questions#
- Link verification asks: are the cited links alive? It checks each cited URL and reports its status. A working link proves the source is reachable, nothing more.
- The analysis asks: do the sources say what the response claims? It reads the response, extracts its factual claims, and tests each claim against the cited evidence.
A response can pass the first layer and fail the second. A live link does not make a claim true.
Link verification#
With Auto track citations on, Cognardo Workbench tracks every link a response cites. A status strip appears under the assistant card in the sidebar's Collect tab, counting the detected links and their check results: verified, failed, and unresolved.
Expand the strip and the link table opens. Each row shows the link, its live status code (a working link reports a success status, such as 200), and how many times it appears in the response. A dead or failing link surfaces here before it reaches your deliverable.
The Auto track citations toggle lives in the pop-up and in settings.
Run the analysis#
- Open the sidebar on the conversation and go to the Collect tab.
- On the ASSISTANT card you want to check, click the Citation and hallucination analysis control. Its tooltip shows the credit cost.
- The card expands with the results:
- The link table, with each cited link and its status.
- The CITATION AND HALLUCINATION ANALYSIS section, with one card per extracted claim.
Read the results#
Every claim card carries the same fields:
- Claim: The factual statement extracted from the response, numbered C1, C2, and so on.
- Verification status: Whether the claim holds against its cited source. A claim the analysis cannot confirm is labelled FACT: VERIFICATION UNCERTAIN.
- Citation: The source backing the claim. A claim with no source is flagged MISSING CITATION.
- Quality: A score for how well the citation supports the claim.
- Tier: A score for the authority of the source itself.
- Issues: What the analysis found wrong or could not complete, in plain terms.
Treat an uncertain or unsourced claim as work: verify it manually or get a citation into the response before the content ships.
The full report#
Click the open control at the top right of the results area and the analysis opens as a standalone report page in a new tab, with the same claim cards at full width, a Last Updated timestamp, and a Back control to return.
FAQ#
Does the analysis use AI credits?+
Yes. It is an AI action, and the control shows its credit cost before you run it. Link verification does not use credits.
Does it work on claude.ai?+
Yes. Both layers behave identically on chatgpt.com and claude.ai.
A link shows a working status. Does that mean the claim is verified?+
No. Link verification only proves the source is reachable. Whether the source supports the claim is what the analysis itself tests.
What does VERIFICATION UNCERTAIN mean?+
The analysis could not confirm the claim against a cited source. That is a prompt to check the claim yourself, not a verdict that it is false.
Where is the report afterwards?+
The report belongs to that conversation turn and opens from the analysis results on its card. Rerunning the analysis on the same turn refreshes it, and the Last Updated timestamp shows which run you are reading.