# Wrong company information in ChatGPT: how to fix Fix wrong company information in ChatGPT and other AI answers: trace each error to its source, correct your pages and records, and report what remains. --- Wrong company information in ChatGPT is fixed at the source, not in the chat window. An AI assistant that states the wrong founding year, an old price, a former executive or a service you never offered is repeating a page, a record or a pattern it found, so the correction is made where that page or record lives and then given time to be picked up again. The work has a fixed order. Find out whether the error comes from a live source or from the model alone, correct the pages and records that carry the fact, publish one authoritative statement of your company facts, let the right crawlers read it, and report what is still wrong. This guide covers each step for the four facts that go wrong most often: dates, services, leadership and pricing. No step below forces an answer to change. What you control is the evidence assistants find, and this guide shows how to make it consistent.
The short version
## Wrong company facts in AI answers come from four kinds of failure Almost every wrong company fact traces to one of four causes: an outdated source, two sources that disagree, no usable source at all, or a different company mistaken for yours. Naming the cause first tells you where to work, because each one is fixed in a different place. An outdated source is the most common. A directory entry, an old press page or an archived price list still says what was true years ago, and an assistant that searches the web will cite it. Disagreeing sources come next: your site says one thing, a profile says another, and the assistant picks one or blends both. When no source states the fact, a model fills the gap with a plausible guess, which is how invented services and wrong executives appear. Entity confusion happens when another company shares your name, or a close variant of it, and the assistant merges the two.
CauseWhat the answer looks likeWhere to fix itSignal it is working
Outdated sourceA fact that was true once, often with a citationThe cited page or record itselfThe citation changes or disappears
Conflicting sourcesAnswers that change between runsEvery page and profile that states the factRuns agree more often
No sourceA confident detail with no citationA clear, crawlable statement of the factThe fact appears with a citation
Entity confusionAnother company's details under your nameDisambiguation on your pages and recordsThe other company's facts drop out
## Test whether the error is live or baked into the model Run the same question several times with web search on and again with it off. If the wrong fact appears only when search is on, a live page is feeding it and you can usually find and fix that page. If it appears with search off, it comes from what the model absorbed in training, and you cannot edit that directly. Do this before touching anything. A model answering without search works from its training, while a search-enabled answer reads current pages and can cite them. Answers vary from run to run, so a single result proves little: five or more runs per question, dated, show whether an error is stable or intermittent.
Log the error before you fix anything
  1. Write the exact promptsUse the questions a customer would ask: who founded the company, what it sells, who leads it, what it costs. Keep the wording identical on every run.
  2. Run each prompt several timesDo it with search on and with search off, in a fresh conversation each time, and record which assistant and which mode you used.
  3. Capture the answer and its citationsSave the wrong statement and every source the assistant links. The cited pages are your work list.
  4. Open each cited pageCheck whether it says the wrong thing, says something ambiguous, or says nothing and the assistant guessed.
  5. Label the causeMark each error as outdated, conflicting, missing or confused with another company, using the table above.
A cited page that does not contain the wrong fact is a finding in itself. It means the assistant misread or over-generalized a source, and the fix is a clearer statement on that page rather than a correction request. ## Fix each type of fact where its error usually starts Dates, services, leadership and pricing fail for different reasons, so each gets a different first move. Start with the fact that costs you most, usually pricing or a service you do not offer.
FactUsual origin of the errorOwned fixOutside fix
Founding dateDirectories that copied a registration or launch dateState the date once, with what it refers toCorrect listings that repeat it
ServicesOld product pages, reseller copy, guessing from the nameA plain list of what you do and do not offerPartner and marketplace descriptions
LeadershipArchived press releases and profile sitesA dated team page with current rolesProfile sites and press pages
PricingCached plans, comparison sites, discontinued tiersA current pricing page with a visible update dateComparison and review listings
### Founding date and company history Dates go wrong because companies have several: incorporation, first trading, product launch, rebrand. State which one you mean in a plain sentence, for example "incorporated in" followed by the year, and use that sentence unchanged on your about page, your facts page and your profiles. An assistant that sees three different years on three pages will pick one at random. ### Services and what you do not offer An assistant guesses services from the company name and the sector. Counter that with a short list of what you offer and, where a wrong service keeps appearing, one sentence saying you do not provide it. Keep that sentence factual and neutral, and do not repeat the false claim in dramatic wording, because a page that dwells on the false claim can become the source of it. ### Leadership and team Former executives linger on old press pages and profile sites that nobody updates. Publish a team page with current roles and a visible date, then work through the profiles that name the former holder. If a person has left, the profile owner is often the only party who can change the record, so contact them with the page URL and the exact line to change. ### Pricing and plans Prices are the least stable fact and the one most often cached. Publish one current pricing page, show when it was last updated, and keep old plan pages from sitting live with a price on them. If you retire a plan, redirect or remove its page and state the change on the pricing page. Comparison sites need their own corrections, since an assistant often prefers a third-party table to your page. ## Publish one company facts page that every other source can match One page that states your core facts, in fixed wording, gives assistants and people a single reference to agree with. Put it on your own domain, link it from the home page and the footer, and treat it as the master copy that every profile is brought into line with. Keep it to the facts that get misquoted: legal name and trading name, what the company is and does, what it does not do, founding or incorporation date and what that date means, headquarters, current leadership, and where to find current pricing. Write each fact as a short declarative sentence, because short sentences are the easiest to lift whole into an answer. Add a date for the last review. Keep it crawlable and in plain text, not an image or a script-only widget. Google's guidance on its AI features says standard search basics still apply, including keeping important content in text form. ## Structured data helps machines read your facts but does not force a correction Organization structured data is markup that labels your company facts so a machine can read them without guessing. Google says it can "help Google better understand your organization's administrative details" on its [Organization structured data page](https://developers.google.com/search/docs/appearance/structured-data/organization). It is worth adding. It is not a lever that overrides what an assistant says. The same page lists no required properties and recommends adding as many relevant ones as apply, among them name, alternate name, address, URL, logo, description and `sameAs`. It also says Google does not promise that features using structured data will appear. Treat the markup as clear labeling, not as a way to command a result. Three practices matter most: - **Match the visible page.** Markup must describe what readers can see. Google's policies say not to mark up content readers cannot see, and a violation can cost a page its rich result eligibility, though Google says this does not affect web ranking. - **Put it where Google says.** Google recommends the home page or a single page describing the organization, such as an about page, and says it is not needed on every page. - **Use `sameAs` honestly.** Schema.org defines it as the URL of a reference page that unambiguously indicates the item's identity, such as an official profile or a Wikidata entry. Point it only at records that are truly yours or truly about you. Validate the markup with Google's Rich Results Test before relying on it. Google states that no special schema is needed to appear in AI Overviews or AI Mode, so do not expect markup alone to change an answer there. ## Correct the entity records that assistants and search engines trust Assistants lean on records about your company that you do not host: a Google Business Profile, a knowledge panel, directory entries, and sometimes a Wikidata item. Correcting these removes the conflicting evidence that makes answers change from run to run. **Google Business Profile.** For a business with a profile, edit it while signed in to the linked account. Google's help page says it reviews changes before they go live and does not state how long that takes, so recheck after submitting rather than assuming the edit took effect. **Knowledge panel.** Google's help page says panel information comes automatically from public information on the web, that descriptions come from data sources and cannot be edited directly, and that Google does not create or delete panels by hand. You can use the Suggest edits option, include public URLs that support the change, and send separate feedback for each item. Anyone can send feedback on a search feature, but the Suggest edits option appears only for a Google Account verified to represent the panel, and verified feedback is prioritized. The page says review usually takes a few days and can take longer. That page is written around the entity you represent rather than a business-specific process, so check how it applies to your panel. **Directories and data aggregators.** These are where outdated dates, addresses and leadership names tend to sit. Correct the ones that rank for your company name first. Each directory has its own process and its own owner, and some require an account you control. **Wikidata.** A Wikidata item can hold structured facts about a company and link to your official site. Whether your company qualifies for an item depends on that project's own inclusion rules. The notability page could not be read for this guide, so check it before creating anything, and never edit an item in a way that serves you if you cannot support it with an independent source. ## Let the right crawlers read the corrected pages An assistant can only repeat a corrected fact it can reach. Check that your robots.txt file and your hosting do not block the crawlers that feed the assistants you care about, then request a recrawl of the changed pages. OpenAI documents separate crawlers. Its [bots page](https://developers.openai.com/docs/bots) says OAI-SearchBot surfaces sites in ChatGPT search, that GPTBot collects content that may be used for training, and that each setting is independent of the others. A site can therefore allow search appearance and still disallow training use. The page also says a robots.txt change for search can take about 24 hours. ChatGPT-User handles actions started by a user, and the page notes robots.txt rules may not apply to it. For Google, the [recrawl guidance](https://developers.google.com/search/docs/crawling-indexing/ask-google-to-recrawl) says to use the URL Inspection tool for a few individual URLs and a sitemap for many. It adds that resubmitting a URL does not speed things up, that crawling can take from a few days to a few weeks, and that a request does not promise inclusion. Submit once and wait. Other assistants have their own crawlers. Where a provider's page for site owners was not read for this guide, do not assume its crawler names or timing. ## Report the error to the provider, and expect little from it Reporting a wrong answer is worth doing once, but it is the weakest tool in this guide. Report, record that you did, and put your effort into the pages and records above. OpenAI's [publishers and developers FAQ](https://help.openai.com/en/articles/12627856-publishers-and-developers-faq) says any public website can appear in ChatGPT search and that blocking OAI-SearchBot can keep a page out of summaries and snippets. As read for this guide, it describes no correction or feedback channel and does not say how fast site changes show. If the assistant offers a feedback option on an answer, use it once and give only the wrong statement and the right one; no OpenAI page read for this guide documents that route or what happens next. For Google, the AI features documentation carries a "Send feedback" link and, in its preview controls troubleshooting, points unresolved issues to the Search Central Help Community. It names no dedicated correction channel for AI features, so treat the link as general feedback. If a false statement is defamatory rather than merely wrong, that is a legal question for your counsel, and the rules differ by country. The routes in this guide do not replace it. This is information, not legal advice. ## A realistic timeline for a correction Plan on days for your own pages, then weeks before assistants consistently repeat the correction, and longer for answers that come from the model alone. These ranges are our experience, not a promise, and none of the provider pages read for this guide states a correction time.
A typical order of work, with usual ranges
  1. Day 1: diagnoseRun and log the prompts, open the cited pages, and label each error by cause. Usually a day or two of work.
  2. Days 2 to 5: fix owned pagesPublish the facts page, correct the about, team and pricing pages, and add Organization markup that matches the visible text.
  3. Week 1 to 2: correct records and listingsEdit the Business Profile, suggest the knowledge panel changes, and contact directory and profile owners with the exact line to change.
  4. Week 2 to 4: request recrawls and re-testSubmit changed URLs once, then repeat the logged prompts weekly with search on and off.
  5. Month 2 onward: monitorKeep testing monthly, or hand it to monthly monitoring. Errors from the model alone may persist, so decide what level of repeat counts as fixed.
Define fixed before you begin. A reasonable test is that most repeated runs of each prompt, over several days, give the right fact. One good run proves nothing, since answers vary. ## A worked scenario with stated assumptions Take an invented software company, used here only as a scenario. Assume it was incorporated in one year but launched its product two years later, retired a mid-tier plan, and had a co-founder leave. An assistant says it was founded in the launch year, quotes the retired plan price, and names the former co-founder as chief executive. The log shows the retired price and the founding year appear only with search on, and both cite a comparison site and an old directory entry. The chief executive error appears with search off too, so it comes from the model as well as a profile page. The plan is ordered by cause. The comparison site and directory are contacted with the current price and the exact founding sentence. The company publishes a facts page that says "incorporated in" the first year and "first product released in" the second, retires the old plan page with a redirect, and adds a dated team page. It edits its Business Profile and sends separate knowledge panel feedback for each wrong item. After weeks of weekly re-tests, the price and founding year are likely to change first, because live sources feed them. The leadership error may last longer because part of it sits in the model, so the company keeps the team page dated and re-tests monthly. Two other habits make things worse: rewriting pages the assistant never cited, and blocking crawlers so the corrected page cannot be read. Fabricated reviews or profiles to outweigh an error break platform rules and invite a second problem. ## The next step is a one-hour audit of what assistants say today List ten questions a customer would ask, run each several times with search on and off across the assistants your customers use, and mark every wrong fact with its cause. That list is your work plan and your baseline for every later test.
If this is happening to you now

If a wrong answer is costing you leads or confusing partners, we can run the audit, trace each error to its source and set the correction order. Request a private review on the audit page.

Reputation Partners handles this work as part of managing what AI assistants say about a company, and ongoing checks are what catch an error early. ## Questions people ask ### Can I make ChatGPT stop saying the wrong thing about my company? You cannot edit what a model has already learned, and you cannot force a search-enabled answer to change. You can correct the pages and records it reads, publish one clear statement of your facts, and re-test. Answers that rely on live sources usually move first, while errors held in the model may last longer. ### Why does AI keep giving old information about my business? Usually because an old page, directory entry or comparison table is still live and gets cited, or because the model learned the older fact and no newer source contradicts it clearly. Find the cited sources, update them, and make the current fact easy to find on a dated page you control. ### Does adding schema markup fix wrong answers? Markup labels your facts for machines, and Google recommends Organization markup on a home or about page. It does not force an assistant to change its answer, and Google says no special markup is needed for AI Overviews. Add it, keep it matched to the visible text, and rely on consistent sources for results. ### Should I block AI crawlers to stop wrong answers? No. Blocking a search crawler can keep your corrected page from being read, so the old fact survives. OpenAI documents its search and training crawlers as independent settings, so you can allow search appearance and still disallow training. Decide on each crawler separately, and read the provider's current page first. ### Who do I contact to report a wrong AI answer? Use the feedback option in the assistant where you saw the error, if it offers one, and keep a record. As read for this guide, OpenAI's publisher FAQ describes no correction channel, and Google's AI features page names none either, only a general send feedback link. Treat reporting as a supplement to fixing your sources, not the main fix. ### How do I know the correction worked? Repeat your logged prompts over several days, with search on and off, and compare to your baseline. A stable majority of runs giving the right fact, with the right source cited, is a reasonable bar. A single correct answer is not proof, because the same question can produce different answers from one run to the next. ### What if another company with a similar name is being confused with mine? State on your facts page who you are, where you are based and what you do, and, where it is true and helpful, which differently named company you are not. Use the same identifying details in markup, profiles and directories so the assistant has distinguishing evidence to work with.