A documented response protocol for correcting brand facts, repairing source evidence and measuring whether inaccurate AI answers continue to appear.
Short answer: answer: Capture the exact answer and citations, classify the error, correct the strongest public sources, use the provider’s feedback or entity controls where available, and retest a fixed query set over time. Repeatedly prompting the system is not a correction method; a material error needs an evidence trail and a monitored propagation plan.
The obvious reaction is to contact “the AI company” and demand a change. Sometimes a feedback route exists, but AI answers can be assembled from search indexes, licensed data, knowledge graphs and multiple web pages. Correcting one output without repairing its source may produce a temporary or query-specific result.
The opposite reaction—publishing dozens of near-identical pages—is worse. Google’s guidance for AI experiences continues to emphasize unique, useful content, crawlability and visible information that matches structured data. More pages do not make a false fact true; they can multiply ambiguity and weaken the canonical evidence.
The AI Misrepresentation Response Protocol
CDM’s AI Misrepresentation Response Protocol has seven steps: capture, classify, trace, repair, report, retest and retain. It treats the output as an observable symptom, not as a database row the brand can directly edit.
Our position is clear: brands should spend more effort repairing authoritative evidence than trying to “train” an answer through prompting. Prompts are useful for reproducing the problem. They do not confer control over retrieval, model updates or other users’ results.
1. Capture the answer as evidence
Record the provider, product surface, account state if relevant, country, language, device, date and time. Save the full query, the complete answer, all visible citations, follow-up context and a screenshot. Open each cited URL and archive the passage apparently supporting the claim.
Repeat the exact query in a clean session and with one neutral paraphrase. Do not run twenty variations and select only the worst answer. The goal is to establish reproducibility, not to create a dramatic anecdote. If the claim concerns safety, financial loss, impersonation, defamation or personal data, route it immediately to the appropriate specialist.
2. Classify what is wrong
Use four error classes:
- Entity error: the system confuses the brand with another entity.
- Attribute error: an owner, location, product feature, price or policy is wrong.
- Temporal error: an old fact is presented as current.
- Interpretive error: sources are accurate but the answer draws an unsupported conclusion.
Also label materiality: low for an inconsequential detail, medium for a fact that affects evaluation, and high for a claim likely to alter a purchase, safety, employment, legal or financial decision. Materiality determines urgency and the breadth of notification; it does not change what is true.
3. Trace the evidence path
Start with cited pages, then inspect the brand’s own source-of-truth page, organization and product markup, profile records, distributor listings, press materials and authoritative third-party databases. Search the exact false phrase in quotation marks. Note conflicting publication dates and whether the erroneous page is indexable.
Google explains that Knowledge Graph facts come from public sources, licensed data and information supplied by content owners. Verified representatives can suggest knowledge-panel changes, while descriptions may require correction at the underlying source.
OpenAI states that public sites can appear in ChatGPT search and that publishers should allow OAI-SearchBot if they want content available for summaries and citations. Those are documented provider statements; neither promises that a corrected page will be selected for a future answer.
4. Repair the authoritative evidence
Update the smallest number of strongest sources. Put the correct fact in visible prose on the canonical page, with a specific “last updated” date, responsible publisher and primary evidence where appropriate. Remove contradictory language from old pages or clearly mark it historical. Make structured data match the visible page; do not use markup to assert something readers cannot see.
If a third-party source is wrong, send a precise correction: quote the erroneous sentence, provide the corrected wording, link to primary evidence and request a dated update. For distributors or directories, update the authoritative account rather than creating another unofficial profile.
This is a CDM recommendation: maintain one fact ledger with the claim, canonical URL, evidence, owner and review date. It turns future correction from detective work into maintenance.
5. Use official reporting routes selectively
Submit feedback on the exact AI answer where the product offers it. For a Google knowledge panel, a verified representative can suggest edits and provide public supporting URLs. Business-profile errors should be handled in the relevant Business Profile controls. Legal-removal routes are for legally cognizable issues, not ordinary disagreement.
The report should identify the exact statement, explain why it is wrong, provide primary evidence and avoid demands about ranking. Keep the case or confirmation number. Do not coordinate employees to flood a feedback mechanism; repeated unsupported reports add volume, not credibility.
6. Retest a fixed query set
Create a small test suite: exact factual question, common paraphrase, comparison question, and one branded query with location or product context. Run it on a defined schedule—daily for urgent high-impact errors, weekly for ordinary corrections—using consistent locations, languages and account conditions as far as practical.
Record answer status as correct, partly correct, wrong or absent; capture citations separately. Bing’s AI Performance in Webmaster Tools, introduced in public preview in 2026, reports citation activity across supported Microsoft AI surfaces. Microsoft cautions that citation totals do not indicate ranking, authority or a page’s role in an answer. Use such data as participation evidence, not a truth score.
7. Retain the correction record
Close the incident only when the authoritative sources are corrected, provider reports are logged, tests show an acceptable state across the defined window, and remaining uncertainty has an owner. Retain before-and-after evidence and the exact test conditions.
An answer may change because of a model update, index refresh or different retrieval path rather than the brand’s intervention. Therefore report the sequence accurately: “the source was corrected on this date and later tests returned the correct fact,” not “our schema forced the model to change.”
Correction evidence pack
Use this asset for every material AI misrepresentation:
| Field | What to record |
|---|---|
| Observation | Provider, surface, query, answer, citations, locale, timestamp |
| Error class | Entity, attribute, temporal or interpretive; materiality |
| Primary truth | Exact corrected statement, source, owner, effective date |
| Source repairs | URL, old text, new text, publication and crawl status |
| Provider reports | Route, case ID, evidence submitted, response |
| Retest suite | Queries, cadence, result, citations and conditions |
| Closure | Accepted state, residual risks, approver and next review |
The pack is deliberately evidence-first. It can support communications, search, legal and product teams without pretending that the brand controls an external answer engine.
Related guides
Frequently asked questions
Can a company force an AI search engine to change an answer?
Usually not through an ordinary content request. A company can correct its sources, use available feedback or entity-management tools and pursue applicable legal routes, but answer generation and source selection remain with the provider. Google, OpenAI and Microsoft publish controls for particular surfaces; none offers a general guarantee that a preferred statement will appear.
The exception is content the company directly controls, which it can correct immediately. Serious unlawful content, impersonation, privacy or safety issues may have specific escalation paths that require counsel or the provider’s formal reporting process.
Should we create a new page to correct every false answer?
No. Strengthen the canonical page that should own the fact unless the subject genuinely deserves a separate resource. Add visible, dated and evidenced language; reconcile conflicting pages; and ensure structured data matches what readers see. A thin correction page created only for machines may add another competing URL and age badly.
The caveat is when the misunderstanding reveals a distinct, recurring user need—for example, a formal product discontinuation or acquisition history—where a dedicated, well-supported page can serve people as well as retrieval systems.
How long does an AI correction take to appear?
There is no reliable universal timeline. The source must be published, crawled or otherwise refreshed, selected during retrieval and reflected in the generated answer. Provider feedback review and knowledge-panel changes can also follow different schedules. Record the source update time and test on a fixed cadence instead of promising a date.
The exception is a surface the brand directly manages, such as its own site or verified profile, where the visible correction may be immediate even though downstream AI answers lag. Escalate high-impact ongoing harm through formal channels rather than waiting passively.
Does schema markup fix incorrect AI answers?
Schema can clarify machine-readable facts, but it is not a correction command or ranking guarantee. Google specifically says structured data should match visible page content. Correct the prose, source evidence and conflicting records first; then use appropriate markup consistently. If the wrong claim comes from a reputable third-party source or an entity collision, adding more markup to one page may not resolve it.
The caveat is that valid structured data can still be valuable for eligible search features and entity clarity, so rejecting schema entirely would be as misguided as treating it as magic.
How do we prove that the correction worked?
Use a predefined query set, test conditions and observation window, then record answers and citations over repeated checks. Success means the material false statement no longer appears under those conditions and authoritative evidence is correct—not that every possible prompt produces identical wording. Track source crawl or index status and provider case responses when available.
The caveat is causality: a changed answer does not prove which intervention caused it. Report correlation honestly and continue periodic tests for high-risk facts because retrieval and model behavior can change.
What if the AI answer cites no source?
Capture it anyway, then investigate the public information environment and use the provider’s feedback route. Search distinctive wording, review entity records and compare other surfaces to find likely propagation paths. Publish the correct fact clearly on the canonical source and remove contradictions you control. Without citations, attribution is weaker, so label any inferred source as an inference rather than fact.
The exception is an urgent harmful statement: use the provider’s safety, privacy or legal channel while the source investigation continues; absence of a citation is not a reason to delay risk containment.
Next decision: How to create a source-of-truth page for a brand, product or creator
Related reading: How to run an AI discovery audit for a brand · Entity clarity for marketers · How to measure brand visibility in AI answers without inventing certainty
Sources and research notes
- Google: How the Knowledge Graph Works, checked 26 September 2026. Supports source and feedback descriptions for Google knowledge information.
- Google Search Central: Succeeding in AI Search, checked 26 September 2026. Supports people-first content, crawlability, preview controls and visible-content/structured-data consistency.
- OpenAI: Publishers and Developers FAQ, checked 26 September 2026. Supports OAI-SearchBot and ChatGPT referral statements; it does not promise inclusion or correction timing.
- Microsoft Bing: AI Performance in Webmaster Tools, checked 26 September 2026. Supports available citation metrics and their stated limitations.
Research limitation: AI products, interfaces and retrieval systems change quickly, and outputs can vary by user and context. The protocol is a CDM operational framework, not evidence of provider-specific causation or a guarantee of removal, ranking or citation.
This article is editorial guidance. Apply the principles in proportion to your market, evidence, and responsibilities.
