An AI is saying something false about your business. How do you fix it?

You do not correct an AI, you correct what it reads. Neither OpenAI nor Google documents any procedure that lets a business get a wrong fact about itself rectified: the thumbs-down button sends a signal rather than triggering a correction, and the only formal request OpenAI offers covers personal data about an individual, not a company's opening hours or prices. Fixing the answer therefore means fixing the sources these engines consult, and that almost always starts with your own website.

This now plays out on two surfaces at once. The same wrong opening hour can sit inside a generative answer that cites no visible page and inside a classic search result that does, and the two are not repaired the same way. Treating them as one problem is the fastest way to spend a month on the wrong fix.

Why can you not just ask ChatGPT to correct itself?

Because no procedure exists for it. OpenAI's help page on searching the web with ChatGPT, updated in October 2026, states that search results and citations can be incomplete, outdated or incorrect, and tells users to open the cited source and check it. The same page explains how to make a website eligible to appear, by allowing the OAI-SearchBot crawler and the search IP addresses OpenAI publishes, and says placement is not guaranteed. It describes no route for a site owner to report that an answer is wrong about them.

OpenAI does run one formal request process, and it is narrower than most owners assume. Its help page on the right to be forgotten, updated in October 2026, describes a personal-data removal request under the GDPR: it requires government-issued identification, it is assessed case by case, and OpenAI writes that it may decline a request where it has a lawful reason to do so. Two lines on that page matter to a business owner. OpenAI says it assesses accuracy from the material supplied with the request and cannot independently investigate disputed facts. And among its balancing factors, OpenAI notes that information closely connected to a person's professional activities may be more likely to remain relevant, which means more likely to stay.

Google's position is the same in different words. Its help page on AI Overviews says the summaries can and will make mistakes, and offers a thumbs-down control followed by a report form. Google writes that this feedback helps it improve AI Overviews. Nowhere does it say that a report corrects a specific answer, or on what timescale.

Where does the wrong fact actually come from?

From your own pages, most of the time. In a study published on December 12, 2024, BrightLocal manually logged the sources ChatGPT displayed across 800 local business searches run in November 2024 in the United States, using two queries across 20 business verticals in 20 cities: business websites accounted for 58% of cited sources, mentions on third-party sites 27%, and directories 15%. On transactional queries, business websites rose to 72%. The study is American and two years old, so the exact proportions do not transfer unchanged to 2026, but the structural point has not moved: the first source of an answer about your business is your business.

The remaining sources can be named one by one. For ChatGPT, OpenAI says it sometimes works with search partners and points to the privacy policies of Microsoft and Shopify, which means a wrong fact sitting in the Bing index has a direct path into a ChatGPT answer. For Google, the Business Profile is the central piece, and Google writes that it gathers information from sources including user reports and licensed content. A wrong opening hour in an AI answer is rarely an invention. It is almost always a wrong opening hour somewhere.

What should you fix first, and in what order?

Fix the sentence, not just the page. A generative engine lifts one to three sentences in isolation, so if your site shows the correct hours on the contact page and stale hours in the footer or in a post from 2023, both are readable and the machine picks for you. The order that works runs from most authoritative source to least.

• Your website: the page in question, but also every leftover copy of the old fact, in the footer, on landing pages, in old posts and in your structured data.
• Your Google Business Profile: address, hours, category, exact name. Google reviews each edit before publishing it and writes that it may not approve a change whose accuracy it cannot confirm.
• Directories and platforms that list you without being asked, starting with the ones already surfacing in answers about your sector.
• Mentions on third-party sites you do not control: an email to the webmaster costs ten minutes and treats the source instead of the symptom.

One detail matters in structured data: if your JSON-LD markup contradicts the visible text on the page, you have manufactured the contradiction the AI will resolve on your behalf. Our article "Structured data for AI: where to start on WordPress" covers that markup and the order to add it in.

How long before the correction shows up?

Days on the Google side, and no guaranteed timescale on the AI side. Google writes that Business Profile edits usually take up to ten minutes to review but can take up to thirty days, and that an approved change can take up to three days to appear in search results. Those numbers are Google's, about Google's own profile. They say nothing about when a generative answer will pick the correction up.

On the AI side there is no commitment on timing because there is no commitment at all. An answer is built from pages fetched at question time when the engine browses, and from the model's own memory when it does not. In the first case the correction can land on the crawler's next visit. In the second it may never land, because a trained model is not updated page by page.

Who can change your business information without telling you?

Any Google Maps user, and your silence counts as consent. Google documents that third parties can suggest an edit to your Business Profile, that it notifies you by email or in-product notification, and that you have four days to accept or reject the suggestion. If you do not respond within that window, Google says it may publish the update automatically where other publicly available information supports it, such as your own website.

So a stale opening hour in your footer is not only read by AI engines. It can serve as the evidence for an edit to your Google profile that you never approved.

When is there nothing to correct?

Three situations are not correction problems, and confusing them with one costs weeks.

• The answer cites no source. There is then no source to fix, there is a gap to fill: a page that states the fact plainly, where the machine went looking and found nothing.
• The answer is one draw, not a finding. The same question asked twice of the same AI does not return the same citations, so seeing an error once does not establish that it repeats. Our article "Asking ChatGPT whether it knows you measures nothing" explains why a single test measures nothing.
• The answer is defamatory or exposes you legally. That is a legal matter rather than a search one, and the person to call is a lawyer, not least because OpenAI writes that it cannot investigate disputed facts itself.

Do you need a tool to watch for this?

Not for a one-off correction. If you know which fact is wrong and where it comes from, an afternoon is enough and it costs nothing: reworking your pages, your Google profile and a couple of directories is manual work nobody should usefully bill you for. Paying only makes sense for the part a human cannot sustain over time, which is establishing that a wrong fact keeps coming back, on which AI engines, and on which questions.

At Biencité the site scan is free and stops at what it can actually measure: what your public files and your home page expose to AI engines, not what AI engines say about you. Question tracking starts at 24 € per month on the Starter plan, with five questions a week, and the Pro plan at 39 € per month tracks fifteen questions a week across five AI engines, ChatGPT, Claude, Perplexity, Gemini and Le Chat. What each check covers is published on our method page. If your problem is a wrong opening hour on a Google profile, no subscription will fix it for you. Start with the profile.