# What Microsoft Copilot says about you: check and correct

What does Copilot say about me? How Microsoft Copilot finds information about people, how to audit it, and which corrections are realistic.

---


When web search is on, Microsoft Copilot answers a question about you by running a web search on your behalf, summarizing what it finds, and listing links to the pages it used. What it says about you is therefore mostly a summary of pages that search can reach, written by a language model that [Microsoft's own transparency note](https://support.microsoft.com/en-US/microsoft-copilot/transparency-note-for-microsoft-copilot) says can make mistakes.

That makes "what does Copilot say about me" a two-part job. First, ask it in a controlled way, from the right product, and record the answer together with its citations. Second, trace every wrong sentence to its cause: a page that is wrong, a page about someone else, or a sentence no page supports. The cause decides the fix, and in most files the fix sits on a web page, not inside Copilot.

Microsoft's pages are described here as of the pages read on October 27, 2026. Menus, labels and policies change, and at least one page read covers an older app version, so check your own app before acting.

<div class="box"><b>The short version</b>
<ul><li>Copilot is several products. A consumer account, a work account and an answer with no web search each behave differently, so test the one the people looking you up are likely to use.</li><li>Run a fixed set of prompts, record each answer with its citations and the date, and repeat the set after every change.</li><li>A citation shows where Copilot looked, not that the sentence is true. Sort each error by cause: wrong source, wrong person, unsupported sentence, or no source.</li><li>Microsoft's pages describe feedback tools for responses, not a removal process for statements about a person. Correcting the source page is the route that usually works.</li><li>Memory and personalization shape your own answers. They do not explain, or fix, what someone else sees.</li></ul></div>

## Copilot is several products, so test the one that matters

The answer depends on which Copilot you ask and whether you are signed in. Microsoft publishes separate documentation for the consumer Copilot used with a personal Microsoft account and for Microsoft 365 Copilot and Copilot Chat used with a work or school account, and the two describe different behavior.

A third case cuts across both. Microsoft's Microsoft 365 Copilot transparency note says Copilot does not require web content or organizational data to produce a response, and that without them it relies only on its language models. An answer about a person built that way has no web page behind it to cite.

<div class="tablewrap"><table><thead><tr><th>Surface</th><th>Account</th><th>Where answers about people come from</th><th>What you can control</th></tr></thead><tbody>
<tr><td><strong>Copilot app and website</strong></td><td>Personal Microsoft account, or signed out</td><td>Web search results it summarizes, with cited links; saved memory if on</td><td>Memory, personalization, history, in-app feedback</td></tr>
<tr><td><strong>Microsoft 365 Copilot and Copilot Chat</strong></td><td>Work or school account</td><td>Web search through the Bing Search service when enabled; organizational data</td><td>User web search toggle; administrator web search policy</td></tr>
<tr><td><strong>Any answer with no web search</strong></td><td>Described for work accounts; consumer pages read do not say</td><td>The language model alone</td><td>Almost nothing at source level; report it and strengthen owned pages</td></tr>
</tbody></table></div>

Test the surface your audience is most likely to use. A recruiter, an investor and a journalist can each meet a different Copilot, and the same name can produce different results in each. If you have both account types, run the audit in both and keep the logs apart.

## Copilot builds answers about people from a web search it writes for you

On the work side, Microsoft's [page on web search in Copilot](https://learn.microsoft.com/en-us/microsoft-365/copilot/manage-public-web-access) says that with web search enabled, Copilot picks out terms where web information would improve the response and sends a generated query to the Bing search service. It says this query is a few words informed by your prompt, not the prompt itself, unless the prompt is very short.

The page's own example shows a person's name becoming the query, so a question about you can put your name into it. This is why a plain name search matters: the pages a search engine returns for your name are the likely raw material. The consumer pages read describe the same search-then-summarize pattern but do not name the search service.

<figure class="keyfacts">
<div class="keyfacts-grid">
<div><b>Bing</b><span>the search service Microsoft's work-account page names as receiving Copilot's generated queries</span></div>
<div><b>A few words</b><span>the size of the generated query, not the full prompt</span></div>
<div><b>Cited</b><span>web-grounded answers list hyperlinked sources below the text</span></div>
</div>
<p class="fig-src">Microsoft's pages on web search in Copilot and the Copilot transparency note, as read for this article. The Bing detail comes from the work-account documentation; the consumer pages read describe the same search-then-summarize pattern without naming the search service.</p>
</figure>

Three consequences follow for a reputation file.

- **Your ranking for your name matters.** A stale directory entry or an old article that ranks well for your name is a likely candidate for the answer, at least where the search service is Bing.
- **Two people with one name can merge.** A short query built from a common name can pull pages about several people.
- **The summary can drift from its sources.** Microsoft's [consumer privacy FAQ](https://support.microsoft.com/topic/27b3a435-8dc9-4b55-9a4b-58eeb9647a7f) says Copilot may misrepresent the information it finds and tells users to verify important facts. The transparency note adds that the models can fabricate content.

Microsoft's Copilot documentation also lets people switch web content off. On the work side the page names a user Web search toggle (Settings, Personalization, Advanced) and an administrator policy called Allow web search in Copilot. Someone looking you up with web search off gets an answer you cannot influence through pages.

## Run an audit you can repeat and compare

Start with a fixed script, not a one-off curiosity check. The point is to run the same prompts after you change something and see whether the answer moved. AI answers vary from run to run because, as Microsoft's transparency note puts it, the models are "probabilistic and can make mistakes".

<figure class="fig">
<figcaption><b>A Copilot audit in seven steps</b></figcaption>
<ol class="fig-steps">
<li><b>Choose the surface and account state</b>Note whether you are signed in, which account type you use, and whether web search and memory are on. Write it at the top of the log.</li>
<li><b>Reduce personal context</b>Turn memory and personalization off for the test, or use a browser profile with no history, so the answers reflect the public web and not your own chats.</li>
<li><b>Run the prompt set</b>Use the five prompt types in the table below, with your exact name as the public uses it, and add a second round with your role and city.</li>
<li><b>Capture each answer in full</b>Save the date, the exact prompt, the whole answer and every cited link. A screenshot is useful as a backup, but copied text is searchable.</li>
<li><b>Repeat each prompt three times</b>A statement that appears in two of three runs is a pattern; one that appears once is noise to watch.</li>
<li><b>Open every citation</b>Check what each linked page actually says about you and whether it is about you at all. Mark each sentence as supported, unsupported or wrong.</li>
<li><b>Sort the errors by cause</b>Use the four causes in the next section. Do not report or request anything until each error has a cause.</li>
</ol>
</figure>

The prompt set is where most audits stay too shallow. A single "Who is [name]?" tells you the headline summary and nothing about the edges.

<div class="tablewrap"><table><thead><tr><th>Prompt type</th><th>Example shape</th><th>What it reveals</th><th>Failure to watch</th></tr></thead><tbody>
<tr><td><strong>Neutral identity</strong></td><td>Who is [name]?</td><td>The default summary and its main sources</td><td>Blending with a namesake</td></tr>
<tr><td><strong>Identity with context</strong></td><td>Who is [name], [role], [city]?</td><td>Whether the right person is found</td><td>Old employer or title</td></tr>
<tr><td><strong>Role fit</strong></td><td>What is [name] known for in [field]?</td><td>The professional story a stranger gets</td><td>Thin or invented achievements</td></tr>
<tr><td><strong>Issue probe</strong></td><td>Has [name] been involved in [issue]?</td><td>Whether a past event dominates</td><td>Overstated or unsupported claims</td></tr>
<tr><td><strong>Source request</strong></td><td>Which pages say this about [name]?</td><td>The citation trail behind the answer</td><td>Links that do not support the text</td></tr>
</tbody></table></div>

Ask the issue probe only about a real issue you already know exists. Asking about events that never happened teaches you nothing.

## Trace each wrong sentence to its cause, using the citations

Every error in a Copilot answer has one of four causes, and each has a different fix. The citations are how you tell them apart.

<aside class="callout">
<span class="callout-label">Rule</span>
<h4>A citation shows where Copilot looked, not that the sentence is true</h4>
<p>Open the link behind every sentence that matters. Microsoft's own pages warn that responses can misrepresent sources, so a cited page is a lead to check, never proof.</p>
</aside>

### The cited page is wrong or outdated

The sentence matches the page, and the page is wrong, old or one-sided. This is the most fixable case. The fix is at the source: update your own pages, ask the publisher to correct or update the item, and use the source's own correction route where it has one. Microsoft's Bing page makes the same point about content on other sites: removing it means contacting the webmaster.

### The cited page is about someone else

The page is accurate for a namesake and Copilot attached it to you. Here nothing needs to be removed. The fix is disambiguation: strengthen pages that carry your exact name with your role, field and location, so a short query has an obvious right answer. Report the blended answer, and keep your own evidence of which person is which.

### The sentence goes beyond its citation

The link is real, but the page does not say what the sentence says. This is the summary drifting from its source. There is nothing to fix at the source, because the source is fine. It is a Copilot output problem: log it, report it with the exact sentence and page, and expect it to vary between runs.

### There is no citation, or the answer is generated without a search

The sentence cites nothing, or the whole answer has no web grounding. You cannot trace it to a page. Report it as a response error, and then do the slower work of making accurate, well-structured pages about you easier to find, so a later search-grounded answer has something correct to cite.

## Report a wrong Copilot statement through the channels Microsoft describes

The pages read describe feedback tools for individual responses. They do not describe a request process for removing a statement about a person from Copilot. Treat reporting as a signal to Microsoft, not a removal right.

For the consumer Copilot, Microsoft's privacy FAQ says to select Give feedback in Settings, or use the flag icon below each response in the mobile app and on the web page. In the mobile app it also describes a long press on a response followed by Report. For Microsoft 365 Copilot, the work-account documentation describes thumbs up and thumbs down buttons that open a feedback pane, and says administrators can review the feedback submitted. On a work account, keep personal detail out of the feedback text.

The consumer transparency note says feedback is reviewed by Microsoft's operations teams and asks users to report harmful or unexpected content that way. It does not promise a correction or a reply, and neither should you expect one.

A useful report is short and checkable:

- The exact prompt and the date.
- The exact sentence that is false, quoted from the answer.
- The correct fact, with a public page that states it.
- Whether the false sentence cites a page, and which one.
- Nothing else: no argument, no threats, no sensitive personal data beyond what the correction needs.

Two other Microsoft pages matter. The [Bing page on reporting a concern](https://support.microsoft.com/en-us/bing/how-to-report-a-concern-or-contact-bing) lets you report URLs in search results, and says Bing limits removal of search results to "a narrow set of circumstances". It also says that even if Bing removes a URL, the content stays at its address until the webmaster takes it down. That page does not mention Copilot, and the pages read do not say whether a Bing change reaches a given Copilot answer. The consumer privacy FAQ does not describe a formal data subject request process and points to the Microsoft Privacy Statement. Whether a data protection right applies to you depends on where you live and on the facts, and it is a separate route from a Copilot feedback report. A lawyer or data protection adviser in your country decides that.

## Correct the source: the sequence that usually works

Because a search-grounded answer is built from pages, changing the pages is the most dependable correction. The routes differ in what they change and what can go wrong.

<div class="tablewrap"><table><thead><tr><th>Route</th><th>What it changes</th><th>Speed</th><th>Risk</th></tr></thead><tbody>
<tr><td><strong>Update your own pages</strong></td><td>The best sources for your name</td><td>Days to publish, weeks to be re-read</td><td>Low, if accurate</td></tr>
<tr><td><strong>Ask a publisher to correct or update</strong></td><td>A third-party source Copilot cites</td><td>Varies widely</td><td>A refusal, or a second article</td></tr>
<tr><td><strong>Report a URL to Bing</strong></td><td>A result in Bing's index, in narrow cases</td><td>Not stated by Microsoft</td><td>Rejection; content stays online</td></tr>
<tr><td><strong>Copilot feedback report</strong></td><td>Nothing directly; it is a signal</td><td>No stated reply</td><td>None, but no result either</td></tr>
</tbody></table></div>

Own pages come first because they are the only route entirely in your hands. State the facts plainly: current role, organization, location, and what you are known for, next to your exact name. If you are building a page that should own your name, the principles on our [personal brand](/personal-brand) page apply, and the point is accuracy and clarity, not volume. Then tell Bing the page changed. [IndexNow](https://www.indexnow.org/) is an open protocol that lets a site owner notify participating search engines of added, updated or deleted URLs, and its site names Microsoft Bing among its supporters. It announces a change and does not tell a search engine how to rank the page, and it only helps if the page is crawlable and the change is real.

Publishers are slower and less predictable. Ask for a correction of a specific factual error, with evidence, and ask for an update rather than deletion when an update is enough. Accurate, lawful reporting about real events is not something we help hide, and a publisher is entitled to say no.

<figure class="fig">
<figcaption><b>A correction sequence over about six weeks</b></figcaption>
<ol class="fig-steps">
<li><b>Day one: log and sort</b>Finish the audit, open every citation, and assign a cause to each error. Usually a single day of work.</li>
<li><b>Days two to seven: act on the sources</b>Publish or update your own pages, send publisher requests with evidence, and send feedback reports with the exact sentences.</li>
<li><b>Weeks two to four: wait for re-reading</b>Pages need to be crawled and re-read before answers can change. In most files this is the slow part, and it varies widely.</li>
<li><b>Weeks four to six: re-run the script</b>Repeat the same prompts three times each and compare with the first log.</li>
</ol>
</figure>

These stages are our experience, not a schedule Microsoft publishes, and none of them is a promise. A steady rhythm of re-testing is what turns a one-off check into monitoring, which is how our [reputation monitoring](/reputation-monitoring) service handles it for clients. For the wider picture of how assistants of all kinds treat a name, our page on [AI search reputation](/ai-search-reputation) sets out the approach.

## Memory and personalization change your answers, not other people's

The pages read describe Copilot memory as something saved from your own conversations, so it can make your audit unreliable. They say nothing that suggests it changes what a stranger sees. If Copilot remembers your job and your projects from earlier chats, an answer to "what do you know about me?" may be about your chat history and not about the public web.

For a personal account, Microsoft's privacy controls page says you can ask Copilot "What do you know about me?" to see what it remembers, and use Delete all Memory to clear it, without deleting conversation history. It says turning off personalization and memory means Copilot forgets its memories of your conversations. That page covers an older version of the Copilot app and says an updated version became available on August 18, 2026, so the menu labels in your app may differ. The consumer privacy FAQ says conversation activity is stored for 18 months by default.

Microsoft's page on managing Copilot Memory in Microsoft 365 Copilot says saving memories is on by default, that saved memories are managed under Chat settings and Personalization in Copilot Chat, and that turning memory off does not delete memories already saved. The page does not say which kind of account it applies to, so check which Copilot you are using before drawing any conclusion.

Run the audit with memory off or in a fresh profile, and log that state. Deleting memory changes your sessions, not what the open web says about you.

## Realistic corrections, with one worked scenario

Copilot answers can be corrected to the extent that the sources can be corrected, and they stay variable even then. The scenario below is invented and generic. It shows the order of operations and what each outcome means.

Assume a consultant whose Copilot answer says she leads a firm she left. The answer cites a business directory listing. Assume that her own site and professional profile state her current role, that the directory page is two years out of date, and that she re-tests weekly for six weeks. The cause is the first one: a cited page that is outdated.

She updates her own bio page so her name, current role and city are stated in plain sentences, asks the directory to update its listing with evidence, notifies Bing of the changed page, and sends a feedback report quoting the exact sentence. After the directory changes and Bing re-reads her own page, the old claim can drop out of the answers. It may still show in some runs while old copies are in circulation, and it may return if the directory is not fixed.

<aside class="callout">
<span class="callout-label">Limit</span>
<h4>Some wrong statements cannot be fixed at the source</h4>
<p>An answer with no web grounding, or a sentence that goes beyond its citation, has no page to correct. Reporting and better owned pages improve the odds, but no one can promise a specific wording.</p>
</aside>

Where a false statement is also a legal wrong, such as a serious false accusation, the question becomes legal and depends on the country, the facts and your evidence. Keep your dated records and take the file to counsel. This guide is information, not legal advice.

## Next step: keep a logged audit, then escalate only what the log supports

Start with the dated log. It shows which errors are fixable at the source, which belong in a feedback report and which need counsel, and it gives you a before-and-after for every change. Reputation Partners runs this kind of audit across several assistants in a private review.

<div class="box hot"><b>If this is happening to you now</b>
<p>If a wrong Copilot answer is affecting a deal, a hire or your family, send us the log you have. We will sort the errors by cause and tell you what is fixable, in what order, and what is not. <a href="/audit">Request a private audit</a>.</p></div>

## Questions people ask

### Can I ask Copilot to forget me?

You can ask it to forget what it saved from your own chats, and you can delete memory and history through the settings Microsoft describes. That does not remove public web pages about you, which are what a search-grounded answer is built from. To change those answers, you change or correct the pages they rely on.

### Why does Copilot give me a different answer from a colleague?

Account type, web search, memory and personalization, and the pages returned that day can all differ between users, and the same prompt can produce different wording. Compare answers only when the test conditions match, and write those conditions in the log so a later difference can be explained.

### Does turning off Copilot web search stop wrong answers about me?

Only in your own sessions. The web content setting that Microsoft's work documentation describes belongs to each user, and administrators can set a policy for their organization. It does not change what someone else sees when they ask about you with web search on, and an answer with no search can still contain errors.

### If Copilot is wrong, will ChatGPT or Gemini be wrong too?

Not necessarily. Each assistant uses its own mix of models, search and sources, and the same wrong page may be used by one and ignored by another. Run the same prompt set on each assistant you care about, log them separately, and treat a shared error as a sign of a shared source.

### Can I take legal action over a false Copilot answer?

Whether a legal route exists depends on your country, the exact words, the harm and the evidence. A lawyer in your jurisdiction decides that. Whatever the answer, a dated log of prompts, answers and citations is the evidence that matters, and it costs nothing to start keeping one now.

### How often should I re-test what Copilot says?

Re-test after every change you make to a source, then at a steady interval such as monthly. Run each prompt a few times, because single answers vary. A live dispute or a hiring season justifies weekly checks for a while, and a quiet name needs less. Log every run with its date and conditions.
