What to do when AI hallucinates facts about your company

When an AI assistant invents a fact about your company, preserve the answer, assess harm, verify the correct fact and trace the likely public evidence before responding. Correct authoritative pages and profiles you control, ask publishers to amend inaccurate records, and use the assistant's feedback or reporting channel when available. Then retest the exact question over time. Do not claim you have "updated the model": businesses generally cannot edit a provider's private training data or internal knowledge directly, and a corrected page may not change every generated answer immediately.

Capture the hallucination before it changes

Save the complete prompt and response, not a cropped sentence. Record provider, product mode, date, locale, account state if relevant, visible citations and links. Note whether the answer was generated once or repeated. Reproduce it with the identical prompt and a small set of natural variations, but do not amplify a defamatory statement publicly merely to document it. The context may reveal that the assistant confused two similarly named companies, inferred an unsupported detail from a real event or repeated an old third-party page.

• Urgent: false claims about safety, legality, fraud, financial status or imminent customer harm.
• High: wrong ownership, product availability, certification, pricing or executive identity.
• Moderate: obsolete locations, policies, feature descriptions or category labels.
• Low: awkward wording, minor omissions or opinions that do not assert a false fact.
• Unclear: claims that require legal, technical or operational verification before classification.

Establish one defensible truth record

Create a short correction brief with the false statement, verified replacement, effective date, evidence and accountable approver. Use primary records for legal identity, official product documentation for specifications and current policy pages for customer terms. If the subject is contested, do not turn a preferred narrative into a "fact." State what is known, qualified and supportable. A permanent public page should answer the underlying question in plain language, link to evidence and show a meaningful update date where freshness affects interpretation.

Correct the source ecosystem in the right order

1. Fix inaccurate owned pages, structured data, feeds, press materials and downloadable documents.
2. Update controlled business profiles and repositories that still publish the wrong fact.
3. Contact cited publishers with the exact sentence, primary evidence and requested correction.
4. Use provider feedback or legal reporting routes when available and proportionate to the harm.
5. Avoid publishing dozens of repetitive rebuttal pages that could create more ambiguity.
6. Log every request, response, source change and verification date for follow-up.

Use the AI readiness checker to verify that the correction page is accessible and understandable. If entity markup contains the wrong name, owner or address, repair it and run the schema validator. Structured data must agree with visible content; it is not a command to an assistant. For an eligible public entity record, the Wikidata guide explains how to work within sourcing and notability rules. Never create promotional entries or manipulate community databases solely to influence AI answers.

Retest without declaring victory too early

Start with an AI visibility scan, then use ModelSaid to preserve recurring tests of the original prompt and related high-risk questions. Record whether the false claim persists, disappears, gains a citation or moves to another wording. Test across providers because their retrieval and update paths differ. A single corrected response does not prove permanent resolution; a single bad run does not prove universal failure. Report sample sizes and examples, and continue monitoring at a cadence proportional to impact.

Build prevention into ordinary publishing

Maintain a governed facts register for company identity, leadership, products, policies and major milestones. Give each fact a canonical URL, owner and update trigger. Redirect retired pages, remove contradictory boilerplate and audit third-party profiles after material changes. Prepare a response template that captures evidence without arguing with the assistant. Add hallucination checks to launch, rebrand and crisis reviews so teams find contradictions before customers do. Train support staff to preserve reports and route them without improvising public rebuttals. Review recurring errors quarterly to identify ambiguous naming, weak documentation or stale distribution partners. The most reliable hallucination response is disciplined information management: make the correct answer easy to verify, reduce contradictory sources, use formal feedback mechanisms responsibly and retain monitoring evidence for the next recurrence.