Hallucinations & Remediation

The hallucination workflow is the part of CiteMetrix that turns observation into action. Citation scans tell you what AI is saying about you. Brand facts tell the system what should be true. The hallucination detector compares the two, and the remediation panel tells you what to do about the gap.

If you only get value from one part of CiteMetrix, this is usually the part. This article walks through how detection works, how to read a hallucination report, and the typical remediation cycle.

How a hallucination gets detected

When a citation scan returns a response that mentions your brand, the hallucination detector takes that response and your full brand facts library and runs them through a fast AI model (Anthropic’s Claude Haiku, for cost and speed). The model’s job is to identify any claims in the AI response that contradict — or appear to contradict — your verified facts.

A flagged hallucination has five attributes:

• Severity — Critical, Moderate, or Minor. Inherited from the brand fact that was contradicted (with the detector’s own judgment factored in for ambiguous cases).
• Type — one of five categories:
• Incorrect fact — the AI stated something verifiably wrong against your facts (e.g., wrong founding year)
• Outdated info — the claim was true once but isn’t current (e.g., naming a former CEO)
• Fabricated claim — the AI invented something that isn’t in your facts at all (e.g., crediting your company with a product you don’t make)
• Misleading context — technically accurate but presented in a way that distorts (e.g., quoting a 5-year-old metric without dating it)
• Critical omission — leaving out something important enough that the response is misleading by absence
• The triggering passage — the exact sentence(s) in the AI response that flagged
• The contradicting fact — which of your brand facts the system compared against
• Detector confidence — Haiku’s own assessment of how sure it is. Low-confidence findings are deprioritized but not hidden, so you can spot-check them.

The remediation panel

Click any hallucination in the dashboard and the right-side panel opens. It shows:

• The full AI response (highlighted at the contradicting passage)
• Which platform produced it, when, and which keyword triggered the scan
• The brand fact it contradicts
• The detector’s classification (severity + type + confidence)
• A remediation recommendation — generated by Anthropic’s Claude Sonnet on demand, the first time anyone opens the panel for that specific hallucination. The recommendation tells you what to fix and where, in plain language. Subsequent views of the same hallucination return the cached recommendation instantly.

The recommendation typically takes 3-5 seconds to generate the first time. After that, it’s cached on the row and free to view forever — fast for you, and important for cost reasons (lazy generation, not eager, is what keeps Anthropic API costs sane at scale).

A typical remediation cycle

Once you’ve got hallucination reports flowing in, the workflow looks like this:

1. Triage. Open the dashboard, sort hallucinations by severity. Critical first. Within Critical, sort by recency — recent contradictions are more impactful because they reflect AI’s current state. The Hallucinations panel supports filters by platform, severity, and date range.

2. Pick one and read the panel. Don’t try to fix five hallucinations in a single sitting. Pick the highest-impact one — typically a Critical-severity contradiction that appears across multiple platforms, or one that contradicts your most prominent positioning. Read the full AI response. Read the triggering passage. Read the brand fact. Confirm in your own judgment that the flag is real. (False positives happen at maybe 5-10% — usually around context-dependent claims or paraphrases the detector misread.)

3. Read the remediation recommendation. It will typically suggest one of:

• Update a specific page on your site — e.g., “Your About page doesn’t currently state the founding year clearly. Add a sentence in the first paragraph: ‘Acme was founded in 2014 in Austin, Texas.'”
• Strengthen your schema markup — e.g., “Your homepage doesn’t have Organization schema with foundingDate. Adding this helps AI crawlers index the correct year.”
• Publish a clarifying piece of content — e.g., “Consider a brief blog post or FAQ entry titled ‘When was Acme founded?’ that directly addresses the question. AI platforms favor explicit Q&A patterns.”
• Address the source of the misinformation — e.g., “An old press release on a third-party site has the wrong founding date. Reach out to request correction or publish a more authoritative version on your own domain.”

4. Make the fix. This is the work that actually moves the needle. CiteMetrix points you at it; you (or your content team, or your agency) execute. The fixes are usually small — a paragraph here, a schema attribute there. The leverage comes from doing them on the right pages, in response to specific AI confusion.

5. Wait, then re-scan. AI platforms refresh their training/retrieval on schedules ranging from days to months. Don’t expect a fix made today to show up in tomorrow’s scans. Most fixes show up in 2-4 weeks for retrieval-based platforms (Perplexity, Grok, Gemini with web access), and 1-3 months for training-based ones (Claude, ChatGPT). CiteMetrix tracks the gap; the trend is what tells you whether your fix took.

6. Mark the hallucination as resolved. Once subsequent scans stop flagging the same contradiction, mark the row resolved. Resolved hallucinations stay in your history (so you can see what you’ve fixed) but drop out of the active queue. If a resolved hallucination resurfaces in a later scan, it’s automatically reopened with a note that it had been resolved previously.

Reading the trend

The dashboard shows a 30-day rolling chart of hallucination volume. Healthy patterns:

• Decreasing total volume — your fixes are landing. AI is converging on a more accurate picture.
• Decreasing Critical volume, stable Minor volume — the highest-impact issues are getting addressed. The trickle of low-severity stuff is normal background noise.
• Sudden spike — usually means either a new content release confused AI, a competitor or news cycle muddied the picture, or you added new brand facts that surfaced existing contradictions. The dashboard lets you click into the spike to see which platform and which fact drove it.

What you don’t want:

• Climbing volume across all severities — your fact library is catching more, but your fixes aren’t taking. Time to look at what you’ve been changing and whether it’s actually reaching the right pages.
• Same hallucinations cycling through resolved → reopened repeatedly — usually means AI is being misled by an external source you haven’t addressed (a third-party article, a wiki, a stale press release). The remediation recommendation should point at it; if it doesn’t, reopen the panel and flag for review.

When the detector is wrong

False positives happen. The detector is an AI model itself, and it can misread context, miss paraphrase, or flag things that are technically inaccurate but immaterial.

If you see a hallucination flag that’s wrong:

• Mark it as a false positive rather than resolved. This signals to the system that the detector misread something, not that you fixed something. Aggregate false-positive data feeds into the detector’s calibration over time.
• Consider whether your brand fact is the issue. Sometimes a fact is too vague or too narrowly worded, leading the detector to over-flag. Editing the fact resolves the underlying noise.

If false positives become a pattern (more than 1 in 10 flags), email eric@citemetrix.com — we’d want to investigate.

What hallucination detection isn’t

• It’s not real-time monitoring. Detection runs as part of citation scans, on the cadence your plan supports. You’ll see hallucinations on the timescale of hours to days, not seconds.
• It’s not a moderation system. The detector doesn’t judge whether AI’s output is appropriate, biased, harmful, or in good taste. It only judges factual contradiction against your verified facts. Other concerns are out of scope.
• It’s not a guarantee of correctness. A clean scan with zero hallucinations means nothing in this scan contradicted your facts. It doesn’t mean AI is universally correct about your brand — only that AI’s specific responses, this run, didn’t cross any of your specific verifications.

Next steps

• If your fact library is thin, add more brand facts — better fact coverage = better detection.
• If you’re seeing too many false positives, review your facts for vagueness or wording issues.
• Set up the PWA to get push notifications when new Critical hallucinations land.
• Configure email digests so the weekly summary lands without you having to log in.