# What ChatGPT says about you, and how to change it

What ChatGPT says about you comes from three layers. Run a logged audit across assistants, diagnose each error, and fix it by the right route.

---


What ChatGPT says about you comes from three layers: what the model absorbed in training, what it retrieves from the web when it searches, and, inside your own account, any memory or custom instructions you have set. Only the first two shape what it can tell other people. To change an answer, run a logged audit, work out which layer produced the error, and fix it by the route that fits: correct the source, use the provider's request channel, use the law, or adjust your own settings.

The question is no longer a curiosity. Employers, investors, journalists and clients ask assistants about people, and an assistant answers in a confident voice whether or not its sources support what it says. A wrong or thin answer can sit in front of someone who never searches the web page that would have corrected it.

This guide gives a repeatable method. It explains what you are actually reading, how to audit several assistants in a way you can compare over time, how to diagnose five common kinds of error, and what each of four routes can and cannot do. It also lists what does not work, because the field is full of promises. It is information, not legal advice.

<div class="box"><b>The short version</b>
<ul><li>An answer about you has three layers: model memory from training, live web retrieval, and your own account's memory, and each layer has a different fix.</li><li>Audit before you act: ask the same questions across several assistants, repeat them, and save dated answers, because the same question can give different answers.</li><li>Diagnose each error as wrong, outdated, thin, confused with someone else, or sourced from a bad page, and fix it at the source where an assistant cites one.</li><li>OpenAI has a request channel for personal data in ChatGPT answers, but it reviews each request case by case and applies only to ChatGPT responses.</li><li>No tactic gives a promise. Re-audit monthly and escalate to a regulator only after a refusal.</li></ul></div>

<figure class="keyfacts">
<div class="keyfacts-grid">
<div><b>No promise</b><span>OpenAI reviews each request to limit information about a person in ChatGPT answers on its own merits</span></div>
<div><b>ChatGPT only</b><span>what an OpenAI removal request changes: responses in ChatGPT, not websites or search engines</span></div>
<div><b>None</b><span>extra technical requirements Google lists for a page to appear as a supporting link in AI Overviews</span></div>
</div>
<p class="fig-src">Sources: OpenAI Help Center, "Right to be forgotten and personal data removal from ChatGPT"; Google Search Central, "AI features and your website", last updated December 2025.</p>
</figure>

## What you are actually reading when ChatGPT answers about you

When an assistant describes you, three different mechanisms can be speaking, and mixing them up is the commonest reason people fix the wrong thing.

The first is model memory. OpenAI says its models are trained on information from the public internet, from partners, and from users, trainers and researchers, with filters meant to remove personal-data aggregators, spam and adult content, and it says it takes steps to limit personal information and trains models to decline requests for private information. Public-figure information appears widely online and may therefore be present. OpenAI also says the same question can produce different answers, because of randomness in generation ([OpenAI Help Center, how ChatGPT and its models are developed](https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed)).

The second is live retrieval. When a question may benefit from the web, ChatGPT can search automatically or on request, and it shows inline citations and a list of sources. An answer that cites pages is, at least in part, a summary of those pages, and an error in it is often an error on one of them.

The third is your own account. ChatGPT can keep saved memories and reference your chat history, which personalize your conversations. OpenAI's help page describes these as part of your account and notes that deleting a chat does not delete a saved memory; it does not describe memory as a way ChatGPT shapes answers to other people. That distinction matters: tidying your memory settings is a privacy step for you, not a way to change what a stranger is told about you.

A fourth fact sits beneath all three. OpenAI has explained that language models produce "plausible but false statements", that training and evaluation can reward guessing over admitting uncertainty, and that arbitrary low-frequency facts, such as a particular person's birthday, are hard to get right (OpenAI, "Why language models hallucinate", September 2025). Its privacy policy tells users not to rely on the factual accuracy of model output. Errors about individuals are therefore an expected failure mode, not a rare bug.

## Run a logged audit before you change anything

An audit has one job: to give you a dated record of what each assistant says, so you can see what is wrong, what is thin, and whether anything you do makes a difference. Without a record, every impression is anecdote.

<figure class="fig">
<figcaption><b>A logged audit of what assistants say about you</b></figcaption>
<ol class="fig-steps">
<li><b>Write eight fixed questions</b>Cover who you are, what you do, what you are known for, where you work or studied, whether there are controversies, what you have published or built, who you are confused with, and whether you would be a safe choice to hire or back.</li>
<li><b>Use fresh sessions</b>Ask in a new chat or a temporary chat, signed out where the product allows, so your own memory and history do not color the answer.</li>
<li><b>Ask several assistants</b>Repeat the questions in ChatGPT with search on and off, in Google's AI features, in Microsoft Copilot, in Claude and in Perplexity, and note the version or mode each time.</li>
<li><b>Repeat each question three times</b>Answers vary between runs, so one answer proves little, and three show whether a claim is stable.</li>
<li><b>Capture the answer and its sources</b>Save the text, the date, the assistant and mode, and any cited pages, because the cited page is usually where the fix lives.</li>
<li><b>Score each answer</b>Mark every claim as correct, outdated, thin, confused with someone else, or sourced from a bad page, and total the scores.</li>
</ol>
</figure>

The log can be a simple table. Four columns are enough.

<div class="tablewrap"><table><thead><tr><th>Question</th><th>Assistant and mode</th><th>What it said</th><th>Verdict and source</th></tr></thead><tbody>
<tr><td><strong>What is she known for?</strong></td><td>ChatGPT, search on, 24 Oct</td><td>Names an old employer as current</td><td>Outdated; cites an old directory page</td></tr>
<tr><td><strong>Any controversies?</strong></td><td>Copilot, 24 Oct</td><td>Attributes a dispute to her</td><td>Confused; the dispute involves a namesake</td></tr>
<tr><td><strong>What has she built?</strong></td><td>Claude, 24 Oct</td><td>Short, generic answer</td><td>Thin; no cited sources</td></tr>
</tbody></table></div>

The rows above are an invented illustration of how to fill the log. The size of the log grows quickly with the design, which is why it needs a template.

<figure class="fig">
<figcaption><b>How many answers an audit produces</b><span>Eight questions; answers per audit design</span></figcaption>
<div class="fig-bars">
<div><span>Four assistants, three runs each</span><i style="width:100%"></i><b>96</b></div>
<div><span>Four assistants, one run each</span><i style="width:33%"></i><b>32</b></div>
<div><span>One assistant, three runs</span><i style="width:25%"></i><b>24</b></div>
<div><span>One assistant, one run</span><i style="width:8%"></i><b>8</b></div>
</div>
<p class="fig-src">Computed as eight questions multiplied by the number of assistants and runs. Widths are each count divided by 96, rounded.</p>
</figure>

A full audit is therefore 96 answers, which is why a spreadsheet beats memory. Run the full version once for a baseline, and a lighter version monthly.

## Diagnose each error by type

Once the log exists, every problem in it falls into one of five types, and each type has its own fix. The table matches the five to the lawful route and states what does not work.

<div class="tablewrap"><table><thead><tr><th>What you found</th><th>Usual cause</th><th>Lawful fix</th><th>What does not work</th></tr></thead><tbody>
<tr><td><strong>Wrong, no source shown</strong></td><td>Model memory or a guess</td><td>Provider request channel, then rectification under data protection law</td><td>Arguing with the chatbot in the chat</td></tr>
<tr><td><strong>Wrong, cites a page</strong></td><td>An error on the cited page</td><td>Correct or remove the page at its source</td><td>Asking the assistant to forget</td></tr>
<tr><td><strong>Outdated</strong></td><td>Old pages dominate the web</td><td>Update your owned pages and profiles, with visible dates</td><td>Waiting for a model refresh</td></tr>
<tr><td><strong>Thin</strong></td><td>Too little citable material</td><td>Publish accurate, consistent pages and third-party profiles</td><td>Special markup or hidden text</td></tr>
<tr><td><strong>Confused with someone else</strong></td><td>A namesake or shared details</td><td>Make your own pages specific, add clear identifiers</td><td>Repeating your name more often</td></tr>
</tbody></table></div>

A sixth type, a false statement of serious wrongdoing, is covered by the legal route below and should be handled urgently with a lawyer. A false claim that you committed a crime is not a thin-content problem.

## Route one: fix the sources

The highest-yield fix is at the source, whenever an assistant cites a page. An answer built on a retrieved page will change once the page changes, and a cited source gives you something concrete to act on.

Start by opening every page the assistant cited and checking whether it says what the answer says. If the page is wrong, ask the publisher for a correction, and use the removal levers in the right order: a platform report where the page breaks a rule, a publisher request where it is inaccurate, a search engine request where a policy applies, and a legal letter only where the content is unlawful. Our [removal service](/remove-negative-content) handles the cases that need more than a polite email.

If the assistant has no cited page, or cites a page that is correct but thin, the fix is to add material. Publish an accurate primary page about yourself, keep the facts identical across your profiles, make dates visible, and earn accurate third-party mentions. Google's documentation says that to appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to appear with a snippet, and that there are no extra technical requirements or special optimizations, no new files and no special schema ([Google Search Central, AI features and your website](https://developers.google.com/search/docs/appearance/ai-features)). The same page describes query fan-out, in which the system runs several related searches, so a wide set of pages can contribute. Ordinary good practice therefore serves both search and AI answers.

When an assistant confuses you with a namesake, the fix is specificity. State your profession, your location and your dates on your primary page, link your profiles to one another, and keep one clear sentence that distinguishes you. Assistants resolve ambiguity from the strength of the signals they find, so give them clear ones.

## Route two: official request channels

Providers run request channels of varying quality, and what each offers should be described as it is.

OpenAI says anyone, with or without an account, can ask it to prevent information about them from appearing in ChatGPT responses, through its privacy portal. It lists grounds that include information that is inaccurate, excessive, irrelevant or no longer appropriate. It reviews each request individually, weighing privacy against freedom of expression and public interest, and the factors it names include public role, accuracy, age of the information, sensitivity, impact, whether the source is official, criminal or misconduct history, and whether a minor is involved. It asks for examples, chat links or web addresses, and identity verification. Its limits are stated plainly: it does not promise removal, it cannot independently investigate disputed facts and relies on what you provide, and removal applies only to ChatGPT responses, not to external sites or search engines ([OpenAI Help Center, personal data removal from ChatGPT](https://help.openai.com/en/articles/20001057-right-to-be-forgotten-and-personal-data-removal-from-chatgpt)). Its privacy policy adds that it considers correction and removal requests based on applicable law and the technical capabilities of its models.

Google's AI Overviews cannot be turned off, and Google's help page says they can make mistakes and invites a thumbs-down and a "Report a problem" with the response. Google's legal reporting tool lists AI Overviews and AI Mode as products, with defamation and other legal reasons for a request, and separate policy reasons such as personal information; the tool notes that legal removals usually apply only in the country where the content is illegal.

Microsoft's Copilot terms say Copilot can make mistakes, direct users to in-app reporting for incorrect content and to a "Report a Concern" page for legal concerns, and state that Microsoft cannot promise its responses will not defame anyone. Anthropic's privacy page for consumer products says that if you have a concern about your personal data in a response from Claude you can contact its support, that its models do not store text like a database, that its ability to identify training data about a person is limited, and that it may decline requests for a lawful reason. We found no dedicated route at Perplexity for reporting an inaccurate answer about a person; its privacy policy states a right to correct inaccurate personal information for California residents, and answers show source links, which points back to route one.

## Route three: data protection law and regulators

If you are in the EU, the EEA or the UK, data protection law gives you rights that apply to what a provider does with your personal data. Article 16 of the GDPR gives a right to have inaccurate personal data rectified without undue delay, and to have incomplete data completed, including by a supplementary statement. The UK has an equivalent. The provider generally has one month to respond, extendable for complex requests, and if you are refused you can complain to your national data protection authority.

Regulators have been asked to apply these rules to AI answers, and the results show how uneven the path is. The privacy group noyb filed a complaint with the Austrian authority in April 2024 about an assistant that kept giving a public figure's date of birth wrongly. Its account of the exchange is that OpenAI said it could only filter or block the name, which would suppress all information about the person, and noyb invoked the GDPR's accuracy, access and rectification provisions ([noyb, complaint page](https://noyb.eu/en/chatgpt-provides-false-information-about-people-and-openai-cant-correct-it)). A further complaint in 2025, reported by TechCrunch, concerned a false claim of serious crime mixed with true personal details, and OpenAI said the version in question had since been enhanced with online search. We could not verify the outcome of either complaint.

Two practical lessons follow. A refusal to correct can end in a block, which means the assistant may decline to talk about you at all, and you should decide whether that is acceptable before asking for it. And a regulator complaint is a slow escalation, appropriate after a provider has refused or ignored a properly made request, not a first step. Where an answer states something false and seriously damaging, such as an invented criminal record, take legal advice promptly, and use the provider's legal reporting channel in parallel.

## Route four: your own settings and crawler controls

Two sets of controls sit in your own hands, and each is narrower than people hope.

Your account controls cover personalization. You can review and delete saved memories, turn memory off, and use temporary chats that do not create memories. These change how ChatGPT treats you in your own conversations. They do not change what it tells anyone else.

Crawler controls cover your own website. OpenAI documents three agents. OAI-SearchBot surfaces sites in ChatGPT's search results, GPTBot collects content that may be used in training, and ChatGPT-User makes visits at a user's request, for which robots.txt rules may not apply. Their settings are independent, and OpenAI says it can take about 24 hours for search to adjust after a robots.txt change ([OpenAI, bots documentation](https://developers.openai.com/api/docs/bots)). Blocking these agents governs what your own site contributes, and it does not edit what third-party sites say about you. If you want accurate pages to inform answers about you, blocking OAI-SearchBot would work against that goal.

## What does not work

A short list of tempting tactics fails or has no evidence behind it.

Paying a firm to "update" ChatGPT promises something no provider has offered. Quoted timelines for correction, such as a fixed number of days, are not backed by any source we found. Special markup, including structured data or FAQ markup added only for assistants, is not described by Google as necessary, and we found no primary evidence that it changes answers elsewhere. Blocking GPTBot to hide facts about you affects your own site only. Deleting your own chat does not change public answers. Keyword repetition on your site helps nothing. And arguing with a chatbot in the chat, or asking it to forget, does not change how it answers other people.

A related caution applies to tracking claims about which sources assistants cite. Some vendors publish studies of the most-cited domains, and those studies measure particular samples of queries, not queries about a person. We did not use any such figures here.

## Re-audit, escalate, repeat

Treat the audit as a routine, because answers change as pages change and as providers update their systems. A realistic cycle looks like this.

<div class="tablewrap"><table><thead><tr><th>Stage</th><th>What you do</th><th>Usually</th><th>Sign it is working</th></tr></thead><tbody>
<tr><td><strong>Baseline</strong></td><td>Full logged audit across assistants</td><td>Week one</td><td>Every claim scored and sourced</td></tr>
<tr><td><strong>Fix sources</strong></td><td>Correct pages, update owned profiles</td><td>Weeks one to four</td><td>Cited pages now say the right thing</td></tr>
<tr><td><strong>Requests</strong></td><td>Provider channels for what remains</td><td>Weeks two to eight</td><td>Responses arrive, with reasons if refused</td></tr>
<tr><td><strong>Recheck</strong></td><td>Lighter audit, same questions</td><td>Every month</td><td>Fewer wrong or confused answers</td></tr>
<tr><td><strong>Escalate</strong></td><td>Regulator or lawyer after a refusal</td><td>When needed</td><td>A formal response on the record</td></tr>
</tbody></table></div>

Keep dated screenshots of every answer you act on, because they are your evidence if you need to escalate. A monthly comparison against the baseline shows whether answers are improving, and our [AI search reputation](/ai-search-reputation) practice runs this same audit as a standing program for clients.

## If an assistant is saying something wrong about you

The first step costs nothing and takes an hour. Write the eight questions, ask them in fresh sessions across several assistants, and save the answers with dates. Then sort every claim into the five types.

<div class="box hot"><b>If this is happening to you now</b>
<p>If an assistant is stating something false or damaging about you, do not argue with it. Reputation Partners can run the audit, trace each claim to its source and tell you plainly which routes apply. Start on our <a href="/audit">audit page</a>.</p></div>

## Questions people ask

### Why does ChatGPT give different answers each time?

Language models generate text with some randomness, and OpenAI says the same question can produce different answers. Search results also change, and an answer that uses live web retrieval can shift as pages change. That is why an audit repeats each question several times and logs the dates.

### Can I delete what ChatGPT knows about me?

Not directly. You can delete your own saved memories and chats, which affect only your account. To limit what appears in answers to others, you can ask OpenAI through its privacy portal, which reviews requests case by case and does not promise removal. It applies to ChatGPT responses only, not to websites.

### Is ChatGPT allowed to say that about me?

It depends on the country and the claim. A false statement of fact that damages your reputation may be defamatory in some places, and data protection law in Europe and the UK gives you a right to have inaccurate personal data corrected. Whether either applies to an assistant's output is still being tested, so take legal advice for serious claims.

### Does turning off memory change what it says about me?

No, not for other people. Memory personalizes your own conversations. OpenAI's pages describe saved memories as part of your account and do not describe them as shaping answers given to others. Turning memory off is a privacy choice for you, not a reputation fix.

### Will updating my website change the answer?

Often, when the answer uses retrieval. If an assistant cites pages, correcting those pages or publishing accurate, indexed, consistent ones can change what it finds. Model memory from training is slower and less predictable to change, so do not rely on a single website update.

### Should I block AI crawlers from my site?

Only if you have a reason. Blocking GPTBot limits training use of your own site, and blocking OAI-SearchBot keeps your site out of ChatGPT search results. Neither edits what other sites say about you, and blocking search can reduce accurate information about you in answers.