Why AI Hallucinations About Your Brand Are a Real Problem
Roughly 12% of brand mentions in major AI assistants contain hallucinated attributes (wrong features, wrong pricing, wrong leadership, wrong product). 73% of B2B buyers trust AI recommendations over traditional ads, so a hallucinated mention can directly affect deals. This guide walks through detection and fix.
Step 1: Set Up Monitoring
Run a recurring prompt suite covering brand name, product names, comparison queries, and category queries across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. Capture the responses verbatim. Tools (Presenc AI, OtterlyAI, Visiblie) automate this; the DIY version is a weekly spreadsheet.
Step 2: Categorise the Hallucination
| Type | Example | Severity |
|---|---|---|
| Wrong product feature | "Acme Headphones support Bluetooth 4.0" (actual: 5.3) | Medium |
| Wrong pricing | "Starts at $149" (actual: $299) | High (affects purchase intent) |
| Wrong leadership / founder | "Founded by Jane Doe in 2018" (actual: Maria Ortiz, 2014) | High (brand credibility) |
| Wrong category | "Acme is a software company" (actual: hardware) | High (positioning) |
| Wrong availability / discontinued | "No longer available" (actual: still selling) | Critical (revenue) |
| Outdated facts | "Acme uses Llama 2" (actual: Llama 4) | Low-Medium |
| Fabricated affiliations | "Acme partnered with [non-real partner]" | High (legal exposure) |
Step 3: Trace the Source
For each hallucination, ask: "Which page on the web does the AI assistant cite when generating this answer?" If the response includes citations (Perplexity, ChatGPT browsing, Google AI Overviews), inspect them directly. If no citations, prompt the assistant to "show me the sources" and inspect what it returns. The hallucination usually traces to one of:
- An outdated brand-owned page (your own pricing page is stale).
- A stale third-party article (review, profile, listicle from 2-3 years ago).
- A confused Wikipedia article (mismatched product / leadership info).
- An incorrect Wikidata entity (wrong founder, wrong founding year).
- A Reddit / forum thread with bad information amplified by frequency.
- A pure model hallucination (no source; the model invented it).
Step 4: Fix the Source
- Brand-owned page is stale. Update the page. Re-publish with the same URL. Add updated date to make freshness obvious.
- Stale third-party article. Contact the publisher with the correction. Many will update if asked politely with a clear correction.
- Wikipedia article is wrong. Provide reliable third-party sources for editors to cite. Do not edit promotional content yourself; it violates Wikipedia's conflict-of-interest policy.
- Wikidata entity is wrong. Fix the entity directly (Wikidata allows brand-owner edits with proper sourcing). Add
sameAsreferences to authoritative sources. - Reddit / forum thread. Cannot edit retroactively. Build counter-evidence with a recent, well-cited post that responds to the misinformation.
- Pure model hallucination. Add a clean, well-cited brand-owned page that gives the model better data to extract from on the next training cycle or retrieval pass.
Step 5: Reinforce with Structured Data
Add or update Schema.org Organization JSON-LD with accurate name, foundingDate, founder, numberOfEmployees, sameAs (Wikipedia, Wikidata, LinkedIn, Twitter). This gives AI assistants an unambiguous machine-readable source. See Schema.org JSON-LD examples.
Step 6: Update Your llms.txt
Make sure your /llms.txt includes the correct brand description with the corrected facts. AI assistants treat llms.txt as a freshness and authority signal.
Step 7: Escalate If Necessary
If the hallucination is material (defamation, false product claims, fabricated partnerships) and persistent after upstream fixes, contact the AI vendor directly. All major AI assistants have feedback mechanisms; persistent escalations through customer success or legal channels work when consumer feedback does not.
Step 8: Monitor for Recurrence
Wait 2-4 weeks after fixing the source, then re-run the prompt suite. If the hallucination persists, the source fix did not propagate (model caching, slow retraining) or there is a deeper source you missed. Iterate.
What Not to Do
- Don't try to "prompt-inject" a correction. Telling the AI assistant the correct answer in a session does not propagate to other users.
- Don't edit Wikipedia promotionally. Violates policy; can result in article protection or deletion. Always go through the editor community with reliable sources.
- Don't ignore Wikidata. Wikidata feeds many AI assistant entity-disambiguation paths and is undermaintained for most brands.
- Don't blanket-blame the AI vendor. Most hallucinations trace to a fixable upstream source, not to the AI model itself.