AI says wrong things about my brand: the correction playbook

Roughly 72% of brands surveyed by Metricus had at least one factual error in AI-generated responses, with an average of 3.4 errors per brand (Metricus). No official brand-correction portal exists at OpenAI, Google, or Anthropic. The fix is indirect: change what authoritative sources say, document the entity, push corrections through platform feedback loops, and in defamatory cases use Article 16 GDPR. Perplexity propagates fixes in days; ChatGPT in weeks to months. There is no faster path.

What is actually breaking

AI hallucinations about brands fall into three categories, and the right response depends on which one you have. The first is factual drift: wrong founding date, outdated pricing, a discontinued product listed as current, or a feature attributed to the wrong tier. These come from stale third-party sources the model retrieved. The second is entity confusion: the AI merges your brand with another company that shares a name, splits CEO history, or attributes a competitor's incident to you. This is a knowledge graph and disambiguation problem. The third is defamatory fabrication: the model invents a lawsuit, a recall, or a criminal allegation that never happened. Each category needs a different correction path. Treat them the same and you waste cycles. Diagnose first, then act.

Why direct correction does not exist

OpenAI's own engineering writing acknowledges that hallucinations are baked into how language models are trained -- they are rewarded for confident answers, not honest "I do not know" responses (OpenAI). That has a practical consequence: there is no support ticket queue at OpenAI, Google, or Anthropic where you submit "your model said X about my company, please change it." OpenAI's content reporting form covers policy violations, copyright, and privacy issues -- not brand inaccuracies (OpenAI). The European NGO noyb filed a GDPR complaint after OpenAI refused to rectify a wrong birth date, with OpenAI arguing it could not surgically correct training data (noyb, TechCrunch). Until that legal pressure forces a portal, your only lever is the corpus the model reads from.

Triage in the first 24 hours

Before you do anything else, prove the hallucination is real, repeatable, and visible to others. A single wrong answer in a long chat does not justify a correction sprint -- context drift inside one session is not the same as a model that consistently misstates facts to new users. Run the same prompt cold (no chat history, signed-out, incognito) on at least three platforms: ChatGPT, Google AI Overviews, and Perplexity. Vary the phrasing across five to ten queries the way a buyer would type them. Capture screenshots and cited sources from every response -- Perplexity and AI Overviews show the URLs they retrieved; ChatGPT's web-enabled answers show citations as well. If the error appears across two or more platforms with at least 30% of prompts, you have a corpus problem. If it shows on one platform only, you have a retrieval or platform-specific problem.

Map the cited sources before you write a single email

The mistake most teams make is rewriting their own About page first. That is the lowest-leverage starting point. Open every cited URL the AI returned and read what it actually says about your brand. In our work indexing AI citations, we see corrections fail when teams update sources the AI never cites and ignore the ones it does. Parse tracks AI visibility across ChatGPT, Google AI Overviews, and Perplexity, and the citation graph regularly surprises operators -- Reddit threads from 2023, archived press releases, comparison-site pages they did not know existed. Build a spreadsheet: cited URL, what it says, who controls it, last updated date, and whether the wrong fact lives there. That spreadsheet is your work order. Anything not on it is a distraction. See which domains AI models cite most for the canonical source-share patterns, and Parse's data on the source domains behind AI answers for the underlying numbers.

Platform timelines: what to expect

The clock is not the same on every model. Perplexity uses live retrieval, so a fix on a high-authority page can show up in days. ChatGPT mixes training data with web search; the search-grounded answer can update in weeks, but the trained-in priors take much longer and may never fully clear. AI Overviews moves with Google's index. Anthropic's Claude refreshes more slowly than the others. The table below is a working estimate for high-authority source updates -- small or low-traffic sites will sit at the slow end of every range.

Fix the sources you control first

Start with what you can change without asking anyone. Update the canonical statements on every property you own -- the homepage, About page, product pages, pricing page, press kit, and every footer link to a "facts" page. Each correction needs a visible "last updated" date and, ideally, JSON-LD that backs the same fact (Organization, Product, FAQPage). Do not bury the correct version in a blog post; put it on the page an AI is most likely to retrieve when answering "what is [brand]." If pricing changed, update the page that has been quoted, not a new pricing post. If a product is discontinued, mark it explicitly on the original product page rather than deleting the URL -- deletion drops the canonical signal and leaves third-party copies as the only source.

Update the third-party pages that are actually being cited

This is the hard, manual part, and it is where most teams give up. Work through the cited-URL spreadsheet in priority order: highest-authority domains first, highest-traffic next, freshest dates last. For Wikipedia, do not edit your own page -- propose changes on the Talk page with citations from independent sources. For Crunchbase, claim the profile and submit edits with proof. For G2, Capterra, and review sites, use the vendor portal; for journalist-written articles, email the reporter a short, factual note with a link to the canonical correction and a public source. Most outlets will not retract, but many will append an editor's note, which is enough for retrieval to pull a different answer. For Reddit threads with persistent wrong information, do not vote-manipulate or argue -- post a single, sourced correction reply from a verified brand account. The correction does not have to win the thread; it has to exist on the page so retrieval can find it. For platform-specific tactics on the citation surface that updates fastest, see how to get cited by Perplexity.

If you want to see which sources shape AI answers about your brand, run a free brand check — it takes a minute.

Make the entity unambiguous

If two companies share a name or your brand has rebranded, the AI is not lying -- it is guessing. The cure is entity disambiguation. Schema.org's sameAs property links your domain to authoritative external IDs (Wikipedia, Wikidata, LinkedIn, Crunchbase, official social profiles), and AI search engines use those links to verify identity (Schema.org). Wikidata is a particularly load-bearing input: Google, ChatGPT, and Perplexity all read from it, and a well-populated Wikidata entry is one of the few signals that crosses platforms (SEO Strategy). Walk through our Wikidata entry guide to publish a clean entity record with the correct founding date, headquarters, products, key personnel, and official website (P856) statements. On your own site, add Organization schema with sameAs pointing to that Wikidata Q-ID, your Wikipedia page, LinkedIn, Crunchbase, and verified social handles. This is the single highest-leverage move when entity confusion is the root cause.

Use the platform feedback loops, but do not rely on them

Submitting feedback inside the AI chat does not "correct" anything in the sense of editing model output. It does generate signal that platform safety and quality teams use to prioritize evaluation. Click the thumbs-down on the wrong response, choose "this is incorrect" or "this is harmful," and in the comment field paste the correct fact, the canonical source URL, and a short note on the impact ("ChatGPT is naming our company in a 2019 lawsuit that involved a different company; here is the original court record"). Repeat from multiple accounts and IPs to broaden the signal. For OpenAI specifically, the privacy reporting form is the right channel for personal-data errors and the EU privacy form is the right channel under GDPR. None of this fixes the model immediately. It does build a paper trail that becomes useful if the issue escalates to legal action.

When to pull the legal levers

If the AI is fabricating a crime, a regulatory action, or a medical/financial harm associated with your brand, this stops being a marketing problem. Article 16 of GDPR gives EU residents a right to rectification of inaccurate personal data, and noyb's complaints have already established that OpenAI's "we cannot correct training data" defense does not satisfy the regulation (noyb, TechCrunch). For executives or directors who are EU residents, file a Data Subject Access Request with each AI vendor, then a rectification request, and escalate to the relevant data protection authority if denied. For trademark misuse -- for example, an AI repeatedly recommending your trademark for a competitor's product -- a takedown notice citing trademark dilution gets faster engagement from legal review than a marketing complaint. In the United States, Section 230 generally protects platforms from liability for user-generated content, but the boundaries for AI-generated output are still being litigated; consult counsel before publicly threatening defamation action.

The 30-day correction cadence

Sequencing matters. The same five fixes done in the wrong order can take three months instead of six weeks. Run the work in waves so each downstream step has fresh, citable evidence to point at.

Track outcomes with a weekly re-run of the same prompt set. If the error persists across three weekly checks after day 30 on Perplexity, you missed a cited source -- go back to the spreadsheet. If it persists on ChatGPT past 8 weeks, the source is likely in the training data and you are now playing the long game with new content velocity. For the diagnostic side, our why AI stopped recommending my brand walks through the same triage logic for visibility loss, and is the AI visibility drop real or noise explains how to separate signal from variance when monitoring weekly.

Prevent the next hallucination from sticking

Correction is reactive. Prevention is what stops a single bad press cycle from echoing for two years. Three habits keep the corpus clean. First, run a monthly audit of the top 30 prompts that decide brand recommendation in your category, log what is cited and what is wrong, and assign owners -- the audit lives in Citation gaps. Second, refresh canonical pages on a monthly cadence with real substance changes, not just date stamps; AI platforms increasingly distinguish refresh from churn. Third, place at least one piece of citable, factual content per quarter on a domain in your top-five citation graph (an interview, a contributed article, a verified review push). The goal is to make the correct version of your brand the most-recently-updated and most-cited version on the web. When that flywheel is running, hallucinations get crowded out instead of fought one at a time. If you want a team to run this correction loop on retainer, Work with us and we can walk through what we have seen work.

FAQ

Does asking ChatGPT to "correct" the answer in the same chat fix it for other users?

No. Corrections inside one chat session only affect that session's context. Other users querying the same prompt get the original answer until either the underlying source is updated or the model itself is refreshed. The platform feedback loops are useful for signaling, but they do not propagate the correction to the model in real time.

How long does Perplexity take to reflect a correction on a high-authority source?

Perplexity uses real-time retrieval, so updates on high-authority pages can appear within days, sometimes inside 24 hours for very fresh changes (Trakkr). Smaller or lower-traffic sources can take weeks. If a fix on a top-tier source has not propagated after three weeks, the issue is usually that the cited URL was not the highest-authority one -- re-check which sources Perplexity retrieves for your prompts.

Can I file a GDPR rectification request against OpenAI for false information about a non-EU company?

Article 16 GDPR applies to personal data of natural persons in the EU/EEA, not to companies as legal entities. If the false information targets a named EU executive or director, that person can file a rectification request as a data subject. For corporate facts unconnected to a named individual, GDPR is the wrong tool -- defamation, trademark, and consumer protection laws in the relevant jurisdiction are usually the better path.

Should I sue OpenAI to get the hallucination removed?

Almost never as a first move. Litigation is slow, expensive, and the legal landscape for AI-generated content is still being written. The faster path is platform feedback plus source-side correction. Reserve legal action for material harm -- a fabricated crime, regulatory action, or sustained refusal to correct after formal rectification requests. Even then, regulatory complaints (data protection authorities) tend to move faster than civil suits.

How do I tell if the wrong answer is from training data versus retrieved sources?

Run the same prompt with web search disabled (or use a model that does not retrieve, like Claude in standalone mode) and compare to the web-enabled answer. If both are wrong in the same way, the error is in training data and will only clear with a model update. If the no-retrieval answer is correct or vague but the retrieved answer is wrong, the error is in a specific source -- and your spreadsheet of cited URLs tells you which one.

Where Parse fits

Parse tracks which prompts return wrong information about your brand, which sources the AI cited to produce it, and how that picture changes week over week across ChatGPT, Google AI Overviews, and Perplexity. The Citations tab surfaces the exact URLs that need correction; weekly monitoring tells you when the fix has actually propagated. If you want a team to run the correction loop end to end -- triage, source-side outreach, entity work, and verification -- Work with us.