How to Get AI Search Models to Stop Citing a Negative Article About You

Last updated: October 2026

A reporter calls you for comment, you decline, the story runs anyway. Two years later the article is still the first thing a client, board member, or recruiter hears about you, except now they are not hearing it from Google. They are hearing it from ChatGPT. Or Claude. Or Gemini. Or Grok. Or Perplexity.

That is the new reputation problem. A single negative article, even one that is factually thin, legally questionable, or three years out of date, gets pulled into the training data or the live retrieval index of every major generative AI model and then repeated back to anyone who asks about you by name. The AI does not just link to the article the way Google does. It summarizes it, often paraphrases it, and presents the paraphrase as fact in a confident two-paragraph answer.

This piece is the practical playbook for getting AI search models to stop doing that. It is written for the two groups who call Digital Crisis Management about this problem every week: individuals who have a story in their results that they cannot shake, and companies whose executives, products, or legal history keep surfacing in chatbot answers to prospects, journalists, and regulators.

The tactics below cover ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Grok (xAI), Perplexity, and Microsoft Copilot. Each one has its own citation mechanism, its own correction channel, and its own blind spots. A real fix usually requires all five tracks running at once.

If you are still deciding whether you need a specialist or whether you can do this yourself, our overview of the top reputation management companies covers who to call and when.

Why AI search models keep surfacing the same negative article

Before you fix it, understand why it keeps happening. There are three mechanisms:

1. Training data. The model learned about you from text that was scraped from the open web before a cutoff date. If the negative article was indexed by Google between then and the cutoff, it almost certainly sat in the training corpus. Most of the major labs scraped Common Crawl, which the Common Crawl Foundation itself documents as one of the largest public web archives used in modern model training (Common Crawl). News sites, court records, SEC filings, and Reddit threads are all over-represented in those corpora.

2. Live retrieval (RAG). Even if the article was published yesterday, modern chatbots do not rely only on training data. ChatGPT with browsing, Google Gemini, Perplexity, Grok, and Copilot all perform live web searches against Google, Bing, or their own index, retrieve a handful of top-ranked pages, and summarize them into the answer. That is called retrieval-augmented generation, and OpenAI documents its live browsing behavior in its own GPT-4 system card and ChatGPT help center. If the negative article ranks on page one of Google for your name, it will be retrieved. Full stop.

3. The “AI about” overviews. Google now shows AI Overviews on name queries, Bing shows Copilot answers, and Perplexity shows a composed summary at the top of the page. These are summaries of the top-ranked pages, which means suppression on classic Google is now a direct lever on what the AI says.

Any plan that only attacks one of those three mechanisms will fail. The article will come back through the one you ignored. The sections below walk through each lever in the right order.

Step 1: Get the article corrected or removed at the source

The single highest-leverage move is deleting or correcting the article itself. If the source URL stops returning the negative claim, every downstream AI, every search engine, and every scraper eventually stops repeating it. There are three paths, in order of likelihood of success:

Correction request to the publisher. Every mainstream news outlet has a corrections policy, and most of them honor factual corrections if you can show the record. The Society of Professional Journalists has published a baseline corrections and clarifications guidance that most newsroom standards editors follow. Lead with the record, not the grievance. A court docket, a settled lawsuit, an exoneration, a retraction from the original source, a date correction, these move the needle. Opinions do not.

Legal removal. A story that is defamatory, that reveals non-public personal information without a public-interest justification, or that violates a court seal is removable through counsel. Section 230 of the Communications Decency Act protects platforms from liability for user content, but the publisher that originally wrote the article enjoys no such shield, and the EFF’s primer on Section 230 scope covers where that line actually falls. We handle this through our content removal practice with outcome-based guarantees.

De-indexing without removal. If the article will not come down, you can sometimes get Google to drop it from search results for your name. Google’s Results about you tool handles personal information removals, and the right-to-be-forgotten channel applies in the EU, the UK, and now California under the Delete Act. The URL stays live on the publisher’s site, but it stops ranking for your name, which starves the retrieval layer.

Correction at the source is slow (weeks to months) but it is the only intervention that moves every AI model at once.

Step 2: Submit removal and correction requests to each AI model directly

Every major AI lab has a formal channel for requesting correction or removal of content about a specific person. Most people do not know these exist. They are not front-page links. They are buried in privacy and support pages, and they are the fastest way to get a single model to stop repeating a specific claim.

ChatGPT (OpenAI)

OpenAI accepts personal data correction and deletion requests through its privacy request portal under GDPR and CCPA. For factual inaccuracies specifically, OpenAI has a separate “report an inaccuracy” workflow inside ChatGPT itself. Submit both. Include: the exact prompt that produced the problematic answer, a screenshot of the answer, the source URLs the model cited, and the documentation that proves the claim is wrong, outdated, or defamatory.

Claude (Anthropic)

Anthropic accepts privacy requests at privacy@anthropic.com and runs its policy under its Usage Policies. Because Claude does not run a persistent live web index the way Gemini or Perplexity does, the main attack surface is the training data itself. A correction request should include the same evidence package as the OpenAI submission, plus a request that the specific URL be excluded from future training runs.

Gemini (Google)

Gemini is tightly coupled to Google Search, which means it inherits Google’s existing removal infrastructure. Three channels matter: the Results about you tool for personal information, the standard legal removal request for defamation and court orders, and Gemini’s in-product “report” button on any answer. The Google removals operate at the search index layer, which means a successful de-indexing starves Gemini’s retrieval the next time someone asks about you.

Grok (xAI)

Grok is the one most people miss. It runs on X (Twitter) data plus live web retrieval, which makes it the fastest AI to pick up new negative posts and the slowest to let them go. xAI publishes its privacy policy and data request process, and corrections should be routed there, with a separate submission to X for any underlying tweets that are driving the citation.

Perplexity

Perplexity is a pure retrieval-plus-summary engine, which is a double-edged sword. It will cite the live top-ranked pages for a query every time, which means suppression on Google and Bing is the fastest lever. Perplexity does accept direct correction requests through its help center and support channels, and it will remove or re-weight specific sources if you can show the source is unreliable, factually wrong, or legally problematic.

Microsoft Copilot and Bing

Copilot uses Bing’s live index. The Bing content removal request is the primary lever. A removal there flushes Copilot’s retrieval for the same query inside days.

Across all six models, keep a correction log: date submitted, who received it, what evidence package was attached, and what the model’s answer looked like before and after. Expect to re-run the audit every 60 days, because retrieval indexes refresh constantly.

Step 3: Attack the retrieval source, not just the model

The dirty secret of chatbot citations is that they are only as good as the top three or four Google and Bing results for your name. If those slots are owned by the negative article, every model will keep citing it, no matter how many correction requests you file.

This is where traditional SERP suppression and Google reputation work still matters more than most AI-era commentary admits. The best suppression campaigns do the following:

• Build or earn five to ten authoritative properties on your name (owned bio pages, LinkedIn, company page, Crunchbase, speaker bios, interview coverage, podcast guest spots, bylined articles in Tier 1 trade press).
• Earn backlinks to those properties from authority domains, which pushes them above the negative article in Google and Bing.
• Register schema markup on your owned pages so Google and Bing understand them as the canonical “about” pages for your identity.
• Add structured FAQ content on your owned pages that answers the exact questions prospects are asking the chatbots, so your pages become the retrievable source.

When the top-ranked pages for your name describe your current record accurately, that is what every retrieval-augmented model will summarize. The negative article drops below the retrieval cut-off and stops getting cited.

For executives and public figures, this is the backbone of what we do in our individual reputation management practice.

Step 4: Build a counterweight the model actually trusts

Suppression gets the article out of the retrieval window. Counterweight content gets the model to say something different about you when asked. These are different jobs, and most firms conflate them.

The AI models are not reading your Twitter bio to form an opinion. They are reading the pages Google and Bing surface, with a strong weighting toward sources that have long-standing authority signals. For individuals and companies both, the counterweight stack usually includes:

Wikipedia and Wikidata. If a notable individual or company does not have a Wikipedia page, create one through a declared, policy-compliant process. If one exists but is thin or inaccurate, use the Talk page and documented corrections. Wikipedia and its machine-readable sibling Wikidata are scraped verbatim into Google’s Knowledge Graph and feed Knowledge Panels directly, as Google’s own Knowledge Panel documentation confirms. Our full approach to this is in Best Reputation Management Companies for Managing Wikipedia.

Tier 1 earned media. A single genuinely-earned feature in Forbes, WSJ, Bloomberg, Fast Company, or a top trade publication outweighs twenty self-published blog posts in every retrieval ranking we have measured. We run this track inside our authoritative media placement service.

Owned-site depth. A real site under your name or your company’s name with three to five pages of substantive content (bio, FAQ, recent work, press, contact) is weighted higher than a one-page splash.

Structured bio pages on vetted databases. Crunchbase, Muck Rack, Clearbit, Bloomberg’s executive profiles, LinkedIn, Wikidata, speaker bureau pages, these are the sources the models trust by default because they are consistently maintained and consistently correct.

Step 5: Watch the knowledge panel and the AI “person” cards

Google’s Knowledge Panel and the equivalent “about this person” / “about this company” cards in Gemini, Perplexity, and increasingly ChatGPT are the single most consequential surface after page one of Google. They are also the one most people never think to monitor.

Google lets a verified person claim and request edits to their Knowledge Panel through the claim-your-panel process. This is how factual errors in the panel (wrong title, wrong company, wrong spouse, wrong birth year, old controversy) get fixed. Once the panel is clean, every downstream AI that scrapes it inherits the correction.

For companies, the equivalent moves are the Google Business Profile, the LinkedIn Company page, Crunchbase, and Wikidata. All four feed the AI “about” cards.

What companies should do differently from individuals

A company facing an AI citation problem has a different threat model than an individual.

For an executive or private individual, one bad article is usually a crisis because there is nothing else in the index. The fix is additive: build counterweight, suppress one URL, correct one claim.

For a company, the negative content is usually one of a dozen signals: a product review, a Glassdoor thread, a lawsuit filing, a regulatory action, a short-seller report, a crisis news cycle. The fix is strategic: decide which signals to attack directly, which to counter with earned coverage, and which to leave alone because the SEO cost of fighting them is higher than the brand cost of letting them exist.

Companies also have one lever individuals do not: the IR page, the newsroom, and the “trust and safety” or “security” page. These are the pages AI models cite when a user asks “is [company] safe?” or “has [company] had any lawsuits?” A well-maintained, substantive version of each of these pages, with current numbers and current language, is a direct counterweight to any crisis citation.

We run this work for companies through our company crisis management practice and for executives through executive crisis management, both with outcome-based guarantees.

The 90-day timeline most engagements actually follow

A realistic AI citation cleanup looks like this:

• Weeks 1 to 2: audit every chatbot’s answer to a battery of ~25 prompts about the subject, catalog every cited URL, grade severity.
• Weeks 2 to 4: file correction and privacy requests at OpenAI, Anthropic, Google, xAI, Perplexity, and Microsoft, in parallel with publisher correction requests on the two or three most-cited source URLs.
• Weeks 3 to 8: ship Wikipedia and Wikidata work, publish Tier 1 earned media, build or deepen owned bio pages, submit schema markup, claim the Knowledge Panel.
• Weeks 6 to 12: measure. Re-run the full prompt battery against every model. Document which answers changed, which still cite the old article, and where to push next.

A clean result at week 12 is: the negative article no longer appears in any AI answer about the subject for the 25 priority prompts, the top three Google results are owned or neutral, and the Knowledge Panel reflects the current record.

A great result is: the AI, when asked, now cites the earned Tier 1 feature or the owned bio page instead.

When to call a specialist

You can run most of Step 1 and Step 2 yourself. The forms are public, the evidence packages are standard, and the correction channels are open to anyone with a legitimate claim.

Step 3, Step 4, and Step 5 are where most self-serve efforts stall. They require ongoing authority-building, Wikipedia and Wikidata editing experience, Tier 1 media relationships, and a monitoring workflow that runs on a 60-day loop against five live AI models at once.

Digital Crisis Management runs this work for individuals and for companies, with outcome-based guarantees on our contracts. We measure success by the AI answer, not by activity. If the chatbots are still citing the same article against you in 90 days, we have not done our job.

If a negative article is still showing up in ChatGPT, Claude, Gemini, Grok, or Perplexity when prospects, clients, or recruiters ask about you by name, reach out for a free consultation.