When AI gets your brand wrong: find it, fix it, prevent it

This guide is for brand, PR, legal and support teams who have found an AI assistant saying something wrong about their company. It covers how to find errors systematically and trace each one to its cause. It then covers which corrections work and how to stop the same error returning. Expect a day to set up checks, and a few weeks for most fixes to show. Fixes that depend on a model’s next training run take longer.

The short version

1. Check what assistants say about you on a schedule, and log each error with its type.
2. Trace each error to its source: a cited page, a confused entity, or the model’s training data.
3. Fix your own pages first, with dates and a visible correction note.
4. Ask third-party sources to correct the record, following their rules.
5. Use the platforms’ feedback channels, knowing what each one can and can’t do.
6. Prevent repeats with one maintained facts page and a change routine.
7. If the error is deliberate, document it and escalate. Don’t answer with fake content.

Fix the source first; platform feedback comes after.

Find the errors

Assistants say themselves that they make mistakes, and sound sure while doing it.

OpenAI says ChatGPT can produce incorrect or misleading outputs and may sound confident even when wrong. Examples include incorrect facts or dates and fabricated quotes or citations. The model may also lack information beyond its training cutoff, or be unable to read a site because of technical issues, paywalls or robots.txt.

Officially documented · OpenAI Help Center

Build a short brand-fact prompt set: who you are, what you sell, prices, locations, leadership, policies, safety and “is [brand] legit?”. Run it monthly in the assistants your customers use (Measuring AI visibility by sampling prompts). Log each error by type, because the type points to the fix.

A 2026 report by the PR agency Wanzhuo Ruiqiao sorts AI errors about brands into four types: invented facts or data, misattributed sources, internal contradictions, and outdated information presented as current. The report describes these from client work and gives no counts.

Observed in practice · observed 2026-05 · 万卓睿桥 (unpublished report, not publicly available; readers can't check it)

Meltwater’s 2026 report gives an example of the outdated type: a brand raises its prices, but AI keeps describing its old price position. The report presents this as a risk, not a measured case.

Hypothesis · Meltwater 融文 (unpublished report, not publicly available; readers can't check it)

Trace each error to its source

Look at what the answer cites. If it cites a page, that page is usually the cause. If it cites nothing, the error probably comes from training data, from mixing you up with a similar name, or from a gap the model filled in.

Google says its AI features use retrieval-augmented generation: they retrieve relevant, up-to-date pages from the Search index and generate a response from them, with links to the supporting pages.

Officially documented · Google Search Central

In a 2026 study of Pilates studios in Hefei, DeepSeek cited complaint and consumer-rights platforms 75 times, 23.9% of all its citations. It also often cited business registry lookups.

Observed in practice · observed 2026-08 · 腾讯新闻（报道合肥图灵象普拉提行业GEO白皮书）

So check the pages you don’t manage too: complaint sites, registries, old directory listings, outdated press articles and encyclopedia entries.

What you see	Likely cause	Where to fix
Wrong fact, cites your old page	Outdated own content	Your site
Wrong fact, cites a third party	Third-party error	That source
Mixed up with another company	Entity confusion	Facts page, structured data, knowledge panel
Old fact, no citation	Training cutoff	Your site and sources; wait for updates
Invented detail, no citation	Gap filled by the model	Publish the missing fact clearly

Most wrong answers have a return address. Find the page the error came from before you complain to the assistant.

Fix your own pages first

Your own site is the one source you fully control, and search-based answers re-read it. For each error:

• Correct the fact on the page, add an “updated” date and, for material errors, a short correction note.
• Fix the same fact everywhere: FAQs, structured data, PDFs, image alt text, old blog posts.
• Remove or redirect retired product and price pages so they stop being retrieved.
• Make sure the corrected pages are indexed and crawlable (AI search access audit).

Correct third-party sources

If the answer cites someone else, ask them to fix it. Each type of source has its own route.

Google says knowledge panel descriptions come from various data sources and can’t be edited. To update one, contact the source and ask for a correction. “Changing the source of the information usually causes the information in the knowledge panel to change as well.”

Officially documented · Google Knowledge Panel Help

• Media articles: ask the editor for a correction, with evidence.
• Directories, marketplaces, review platforms: update your listing, or use their business-owner correction process.
• Complaint platforms: respond and resolve the complaint where possible. An answered complaint reads differently from an ignored one.
• Encyclopedias: follow their sourcing and conflict rules.

Baidu Baike requires public media reports or business registration records for key company facts such as founding date, funding and honours. Wikipedia asks editors with a conflict of interest to propose changes on the article’s talk page rather than edit directly. In articles about living people, very obvious errors may be fixed quickly, including by the subject.

Officially documented · 百度百科：参考资料 · Wikipedia:Conflict of interest

Use the platforms’ feedback channels

Feedback channels exist, but they are for reporting, not editing. Use them after fixing the source, and include links to the corrected pages.

For an AI Overview that is “unhelpful, inaccurate, biased, or otherwise problematic”, Google asks users to click the thumbs-down, choose Report a problem and add details. The feedback includes the most recent search query and its results.

Officially documented · Google Search Help

An official representative can claim a knowledge panel by signing in to one of the official sites or profiles listed, such as YouTube or Search Console. Not all panels are claimable. Anyone can use Feedback on a panel to suggest a change, with links to pages that verify it.

Officially documented · Google Knowledge Panel Help

Google says it reviews feedback from verified users within a few days, and sends a confirmation email with a resolution update.

Officially documented · Google Knowledge Panel Help

Google may remove Knowledge Graph information that is demonstrably false or outdated, as shown by legal documents, expert consensus or other reputable primary sources. “We may decline to act on facts that are reasonably disputed or lack demonstrative evidence.”

Officially documented · Google Knowledge Panel Help

OpenAI’s privacy policy says individuals who find factually inaccurate information about themselves in ChatGPT output can request a correction or removal through its privacy portal. OpenAI will consider the request based on applicable law and the technical capabilities of its models.

Officially documented · OpenAI

That route is for personal data about individuals, such as a founder. We found no documented OpenAI channel for correcting facts about a company, and none for Perplexity or Microsoft Copilot beyond in-product feedback. For Chinese assistants, we also found no public documentation of a brand-correction process.

The same PR agency report describes client cases on DeepSeek and other Chinese assistants that combined monitoring, publishing authoritative content and in-platform feedback to correct errors. It reports results such as more citations but discloses no method, and doesn’t separate the effect of feedback from the other steps.

Observed in practice · observed 2026-05 · 万卓睿桥 (unpublished report, not publicly available; readers can't check it)

Feedback buttons report a problem. They don’t edit the answer. The fix that lasts is a corrected source.

Prevent repeats

NBD’s 2026 GEO Red Book recommends a correction routine for content that has caused wrong AI citations. Correct the source page, with an update date and a correction note. Fix structured data, FAQs, image alt text and PDFs. Request re-crawling or send feedback to search engines and platforms. Notify media, partners and customers who used the content. It also suggests tracking a “fact error rate”: the share of sampled answers containing a factual error about the company. These are recommendations, not tested results.

Hypothesis · 每经AI智库 (unpublished report, not publicly available; readers can't check it)

Put this into practice:

• One facts page that states your core facts plainly and is the single place you update first (New brands: why AI doesn't mention you yet, and what to do first).
• Organization structured data that matches it (Organization structured data).
• A change routine. When prices, products, leadership or addresses change, update the facts page and listings the same week.
• Monthly checks with the brand-fact prompt set, and a log of errors found and fixed.

What not to do

• Flood the web with rebuttal content. Mass-produced articles are what regulators call poisoning, even when your cause is just (GEO regulation in China).
• Edit encyclopedias without disclosure, or ask an agency to.
• Threaten the assistant’s operator before fixing the source. Without a corrected source, there is little they can point to.
• Assume one fix holds. Re-check after each model or product update.

If the errors look deliberate, treat them as an attack, not a hallucination.

The China AI Industry Alliance’s 2026 report describes GEO abuse that targets competitors. It includes mass-producing false negative news, inventing safety problems, and altering third-party encyclopedia and Q&A content to plant misleading facts.

Observed in practice · observed 2026-05 · 中国人工智能产业发展联盟（AIIA） (unpublished report, not publicly available; readers can't check it)

Collect evidence (answers, cited URLs, dates), report the source content to the hosting platform, and involve legal counsel.

How you’ll know it’s working

• Error log trend. Fewer open errors month over month, by type.
• Fact error rate. The share of brand-fact answers that contain an error, per assistant (Measuring AI visibility by sampling prompts).
• Citations. Are answers now citing your corrected pages instead of the old sources?
• Time to fix. How long each error type takes from detection to a correct answer. Use it to set expectations internally.

Evidence at a glance

Recommendation	Evidence
Assistants make confident factual errors	Official (OpenAI)
Search-based answers re-read current pages, so fix the source	Official (Google)
Knowledge panel text changes when the source changes	Official (Google)
AI Overview and knowledge panel feedback channels	Official (Google)
Personal-data correction requests to OpenAI	Official (individuals only)
No documented company-correction channel for most assistants	Our search of public documentation
Correction routine and fact error rate	Hypothesis (industry recommendation)
Unmanaged third-party pages get cited	Observed
Deliberate attacks on brands exist	Observed (industry report)