ChatGPT vs Perplexity vs Gemini sources: what differs

ChatGPT, Perplexity and Gemini do not show the same thing when someone asks about you. Perplexity puts numbered citations on every answer. ChatGPT cites when it searches, and answers from memory when it does not. Google spreads its answers over several surfaces, each with its own thumbs-down and its own rules. The sources differ on screen, the feedback routes differ, and the audit you run has to follow those differences.

This guide compares them on four questions: where an answer is said to come from, how sources are displayed, how to report an error, and what each assistant remembers while you test it. It compares only what each provider's own help and privacy pages state. Those pages were read for this guide, are named in the text, and describe the product as of the day they were read. Most show no date, or only a relative one such as "updated 2 months ago", so dates are given in the text only where a page shows one.

It does not rank the assistants, and it does not claim to know how any of them picks one page over another. No page we read explains that, and several say plainly that they do not describe it. What the pages do give is enough to build a side-by-side table and a routine you can repeat every quarter.

Where each provider says its answers come from

Every provider describes two origins: what the model learned in training, and what it fetches from the web when it searches. The pages differ in how much they say about each, and in whether the search is always on.

Related readAI Answers About Job Candidates: Audit and Fix

OpenAI's page on whether ChatGPT tells the truth says ChatGPT "does not search the web unless specific tools are enabled", and adds that these tools are enabled for all models by default. Without search, answers come from training. With search or deep research, the model can access and cite real-time web sources. A separate OpenAI page says training uses openly accessible internet information, third-party partner data, and information users and trainers provide, and that training data may incidentally include personal information.

Perplexity's help page says it uses advanced AI to search the internet in real time, then compiles the most relevant insights into an answer. It describes Deep Research as running dozens of searches and reading hundreds of sources, and says Pro subscribers can choose the underlying model.

Google describes two grounded features. Its Search Central page on AI features says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources, a technique it calls query fan-out, and that the two may use different models and techniques, so the responses and links vary. The Gemini app is covered separately by Google's Gemini privacy page, which says language-model experiences, Gemini Apps included, can hallucinate and present inaccurate information as factual.

Microsoft's transparency note (dated August 18, 2026 on the page) says Copilot is grounded in web search results and uses the top results to generate a summary, but only in certain conversations where users seek information. Anthropic's web search page says Claude invokes a search tool to ground its responses in content from the live web, when the question benefits from current information.

Related readAI answers about doctors and lawyers: what to fix

The many cells marked as not described are the finding, not a gap in the reading.

How the sources appear on screen

The visible difference is how much evidence you get with each sentence. That decides how quickly you can trace a claim to a page, and therefore how fast a fix can start.

Perplexity says each answer includes numbered citations linking to the original sources. A separate page on source labels says it attaches small shield-icon labels to some cited domains: Government, Academic and Trusted. The page says labels describe the type of website, that Perplexity rates the whole website rather than each page, and that most domains have no label, which does not mean low quality. A label is not a verdict on the page that mentions you.

OpenAI's ChatGPT search page, in the version we read, says responses that use search contain inline citations, that on desktop web you can hover over a citation for more detail, and that a Sources button at the end of the response lists all cited sources and other relevant links. That version still described access in terms of paid plans and a waitlist, so it may lag the product; check it before relying on a detail.

Google's AI Mode page says to click the links for AI sources and to try other Google Search results too. It also says AI Mode shows a set of web links instead of an AI answer when it does not have high enough confidence. The AI Overviews help page says the snapshot includes links to supporting information from the web.

Related readWrong company information in ChatGPT: how to fix

Copilot's transparency note says the output contains hyperlinked citations listed below the text. Claude's web search page says every response includes citations, direct links and, where appropriate, relevant quotes.

A citation shows that a page sat next to a sentence, not that the page says it. OpenAI's accuracy page lists fabricated quotes, studies, citations and references to non-existent sources as a known type of error and advises checking sources by visiting the links directly. Copilot's note says web-grounded responses might misrepresent content. Open every link.

What the pages do not say about choosing sources

The weakest point in most comparisons you will read is the claim that one assistant prefers one kind of source and another prefers a different kind. The provider pages read for this guide do not support that kind of claim.

Perplexity describes interpreting the question, searching and compiling. Google describes fan-out and says there are no additional requirements to appear in AI Overviews or AI Mode beyond being indexed and eligible to be shown in Search with a snippet, and that meeting every requirement does not ensure crawling, indexing or serving. Microsoft says Copilot "centers its response on high-ranking content from the web", which tells you search ranking matters to Copilot without telling you the weighting. OpenAI's page names a search provider that queries may be shared with, in the version we read, and does not describe a ranking method.

This matters for strategy. If sources are chosen in ways nobody outside the provider can see, the dependable lever is the pages themselves. A source you can edit, correct, update or outrank with accurate content changes what is available to every engine at once.

Related readGoogle AI Mode reputation: audit what it says

How feedback and corrections differ by provider

Each provider offers a quick control on the response. Only some publish a distinct route for personal data or a legal report. The routes below are as the pages read describe them, and none of the pages promises a result.

OpenAI publishes a page on the right to be forgotten and personal data removal from ChatGPT. It says to use the Privacy Portal, choose "Remove personal data from ChatGPT responses", give specific examples with chat links and referenced URLs, name the data type, and explain why the data is inaccurate, excessive, irrelevant or outdated. It says government ID or alternative proof will be required, and that each request is assessed case by case against interests such as freedom of expression and the public interest. It also says removal prevents the information from appearing in ChatGPT responses and does not remove it from external websites or search engines.

Perplexity's page on reporting incorrect answers says to use the flag icon below the answer, or create a support ticket, with the URL of the query and a description of the error and the expected result. It does not say what happens after you submit.

Google's AI Overviews page says thumbs up and thumbs down icons appear at the bottom of each overview, that negative feedback goes through "Report a problem", and that the feedback includes your most recent search query and its results. For the Gemini app, the Gemini Apps privacy hub says that, where applicable privacy law provides these rights, you can object to processing of your personal data or ask for inaccurate personal data in responses to be corrected, through the Help Center or through "Report legal issue" under a response in the app.

Related readWhat Siri, Alexa and Google Assistant say about you

Anthropic's page on reporting content describes a content reporting form for output that violates its usage policies or local laws, and says it does not address inaccurate information about a person directly. Microsoft's note says to report harmful or unexpected content through feedback, and describes no separate personal-data lane.

A practical reading: a quick control is for wording problems that are not personal data. A personal-data route is for statements about an identifiable person, where the provider's page says it exists. A legal route is for content you believe is unlawful. Each asks for different evidence, so filing in the wrong order wastes time.

Memory and history change what you see while you test

An audit is only useful if the answers reflect what a stranger would see. Several assistants say they carry context or history, which can distort a test run on your own account.

Perplexity's help page says it remembers the context of previous queries. Google's AI Mode page says personalization requires you to be 18 or older with history and personalized recommendations enabled, and that Personal Intelligence uses previous searches and activity saved in your Search Services History. It adds that you can correct the system with follow-ups and ask it to remember or forget details. Microsoft's note says Copilot can store specific facts users ask it to remember and can delete them when asked to forget, and that personalization can be turned off in settings.

Google's Gemini privacy hub, dated October 23, 2026 on the page, is the most detailed on retention. It says activity is auto-deleted after 18 months by default, that a subset of chats are reviewed by human reviewers, that reviewed chats are kept for up to three years even if you delete your activity, and that temporary chats and chats with Keep Activity off are kept for 72 hours.

Related readWhat Microsoft Copilot says about you: check and correct

Three consequences follow. Test in a fresh thread for every question, using a temporary or logged-out mode where one exists. Write the surface and settings next to each answer. And do not paste private details about yourself into a test prompt, because your own chat history is data you may not want reviewed.

What the differences mean for a person or brand checking their name

Three decisions follow from the comparison.

The first is the order of work. Where citations are visible, read them first. Where an assistant answers without a citation, treat the sentence as unsupported until you find the page behind it by searching the claim yourself. A claim with no traceable page cannot be fixed at its source. It can only be reported or outweighed by an accurate, better-placed page.

The second is the choice of route. Wrong personal data about you goes to a personal-data route where one is documented. An unlawful statement goes to a legal route. A clumsy sentence that is not false goes to the quick control, or is left alone. Using the heaviest route for the lightest problem slows everything.

The third is how often to check. A quarterly round, with an extra run after any event, is a sensible default from experience. No provider documents a refresh schedule.

Do not file reports to make accurate, lawful, public-interest information disappear. OpenAI's removal page says requests are weighed against freedom of expression and the public interest, and such a request can fail. If something true is unflattering, the lawful lever is better accurate content on pages you control.

Related readChatGPT wrong information about my product: how to fix

A repeatable routine for checking one name across assistants

The routine below usually takes about two hours for the first round and less afterwards.

1. Fix the question setWrite 8 to 12 questions a stranger would ask, including one naming your employer or product. Keep the wording identical every round.
2. Choose the surfacesList the surfaces you will test: ChatGPT with search, Perplexity, a Google results page with AI Overviews, AI Mode, the Gemini app, and any others your audience uses. Write down the account, device and settings for each.
3. Limit memory effectsUse a fresh thread per question, and a temporary or logged-out mode where one exists. Turn off personalization if the product lets you, and say so in the log.
4. Run each question at least three timesSpace runs across more than one day.
5. Log text, date, surface and citationsCopy the full answer, the date, the surface, the settings and every cited page. If there is no citation, write "none".
6. Open every cited pageCheck that the page says what the sentence says. Mark each claim as correct, outdated, wrong, unsupported or about someone else.
7. Pick the route per claimWording problem: quick control. Wrong personal data: personal-data route. Unlawful content: legal route. Wrong or old source: contact the owner of the page or update your own.
8. Fix sources before reportsEdit what you own, request corrections from publishers, and publish accurate pages where the record is thin. Then rerun the same questions after a few weeks.

A difference between two surfaces is information. A difference between two runs on one surface is noise until it repeats.

A four-week cycle from first audit to first check

This is how a first cycle usually runs in our experience. Pages update on their own timetables, so treat the weeks as a plan, not a forecast.

Some claims will not move in four weeks. A publisher may not reply, and a provider may decline a request. A dated log shows what you tried and lets you decide calmly what to escalate.

An invented scenario: one outdated profile, three different answers

This scenario is invented, with generic parties, and the assumptions are stated so you can see how the comparison changes the work.

Assume an architect whose 2019 trade-magazine profile names a firm she left in 2022. She runs the question set on three surfaces. Assume the Perplexity answer cites the magazine page for the old firm name, the Google AI Overview links to the same magazine and to her current firm, and the ChatGPT answer names the old firm with no citation.

The Perplexity result is the easiest to act on. The citation points to a page with a date and an editor. She writes to the magazine, asks for a dated update note, and updates her own biography page to state the current firm and the move. The Google result points to two pages, one outdated and one current, so improving the current one and the magazine page both matter. The ChatGPT result has no citation, so she searches the old firm name next to her own, finds the magazine page as the probable origin, and logs the answer as unsupported until the page changes.

Related readAI due diligence on executives: what to trust

She files nothing with any provider in week one. The old firm name is outdated, not harmful or unlawful, so she corrects the sources and reruns. If after four weeks one assistant still states the old firm with no source, she uses that provider's quick control, citing the exact answer and the correct fact. Had the statement been an invented claim about her conduct, she would move to the personal-data route where the provider documents one.

Where this comparison stops

Several neighbors of this topic deserve their own treatment, so each gets one paragraph here. What ChatGPT says about a named person, what Perplexity's citation list shows about you, and how Google's several products speak at once are each worth a dedicated audit. Hallucinations about people, meaning fluent and specific false statements that no source supports, are a separate subject from errors copied out of a page. Same-name confusion, where an assistant mixes you with another person, needs disambiguation work on your own pages. This guide only tells you where to look: in the cited page, if there is one.

Next step: run one round on two surfaces this week

Pick the two assistants your clients, investors or recruiters are most likely to use, write ten questions, run each three times, and open every citation. That one round shows which differences in this guide affect you.

Reputation Partners usually starts a file with a log like the one above, because it separates what the sources say from what the assistants add.

Questions people ask

Which assistant shows the most sources?

The pages read do not support a ranking, and counting links across products is not a fair measure. Perplexity says each answer carries numbered citations, and Claude says every web-search response includes citations. ChatGPT, Copilot and Google describe links or citations on search-grounded answers. Judge each assistant by whether you can trace the claims about you to a page.

Related readAI defamation: your legal options when AI lies

Does a citation mean the answer is correct?

No. A citation shows which page sat next to a sentence, not that the page supports it. OpenAI's accuracy page lists fabricated citations as a known error, and Copilot's note says even web-grounded answers might misrepresent content. Open the link and compare the page with the sentence before you trust either one.

Can I stop an assistant using my website?

Providers document crawler controls for site owners, and Google says robots.txt controls crawling for Search while snippet controls such as nosnippet limit what is shown. Anthropic's page says removing content is the best way to keep it out of Claude web search outputs. Blocking has costs, including less visibility, so decide with your goals in mind.

Should I use thumbs down or the privacy form?

Use the quick control for a wrong or poorly worded answer that does not involve personal data. Use the privacy route, where a provider publishes one, when the answer states personal data about you that is inaccurate, outdated or excessive. OpenAI says requests are assessed case by case and require proof of identity, so prepare evidence first.

Will my test questions be stored?

Possibly. Google says Gemini Apps activity is saved when Keep Activity is on and can be reviewed by humans, with reviewed chats kept for up to three years. Perplexity says it remembers the context of previous queries within use. Check each provider's privacy page, avoid private details in test prompts, and use temporary modes where they exist.

How often should I repeat the check?

A quarterly round, plus an extra run after any event that touches your name, is a reasonable default from our experience. No provider documents a refresh schedule. Because results vary between runs, compare rounds on the same questions and settings and act on claims that repeat, not on a single answer.