You correct what AI says about your business by correcting the records it reads, not the answer it gave you. In practice that means four surfaces: your Google Business Profile, the listings and directories each engine licenses, the pages on your own site, and the independent sources that corroborate both. There is no correction desk inside the assistant. Every route the platforms publish points at the sources instead.
Most owners try the obvious thing first, which is arguing. You type: that is wrong, we moved in 2023, and we stopped doing that work years ago. The assistant apologizes, gracefully, and agrees with every word. Then you open a fresh window the next morning and the old version is back, in the same confident voice, for a buyer who will never see your correction. You were talking to the reader. The library never heard you.
That is not a verdict on your business, and it is not you failing to find the right button. There is no button. There is a set of records instead, each with its own owner, its own form and its own review queue. You can edit the ones you control outright, and request changes to the rest.
Here is the map. What a correction actually is, the three separate Google front doors and what each one reaches, and what ChatGPT can and cannot get to. Then why a single fix arrives on several different schedules, and the small number of cases where the first call should be a lawyer rather than a login. If you have not yet worked out which kind of wrong you have, start with the diagnosis. The four ordinary causes behind incorrect or outdated AI answers about your company are stale model memory, stale sources, mistaken identity and records that disagree. They need four different repairs. This guide is the repair half.
What a Correction Actually Is
Two different things get called "what AI says about my business," and they take completely different work.
The first is model memory: whatever a model absorbed from the open web before its training cutoff. That material is baked in. Nothing in the Google and OpenAI documentation cited here offers a business a route into a model's learned parameters, and anyone offering to reach inside a model on your behalf is selling something nobody has published a way to do. What does shift memory over time is the public web itself: what it says about you before the next model gets trained. That is slow, indirect and genuinely worth doing. It is never a same-week fix.
The second is the retrieved record: the listings, profiles, pages and licensed feeds an assistant can reach at the moment somebody asks. That half is editable, often by you, usually today, and it is where every practical correction lands. We mapped which assistant looks where in a companion piece on where ChatGPT gets information about your business.
There is a third thing worth ruling out early, because it wastes a lot of afternoons. Where an assistant keeps a memory of your own chats, that memory is account-specific. Correcting it there changes what you see, and nothing about what a stranger in another city gets told. So the useful mental model is not a conversation at all. A correction is a filing. You are not persuading a machine. You are changing the evidence it will find next time it looks, and the evidence has addresses.
The Google Side: Three Front Doors
Google is the surface with the most documented correction routes, and owners routinely use the wrong one, then conclude that nothing works. Three of them cover most business-information problems, and they govern different objects.
One. Your Business Profile. This is the listing behind Google's local answers and Maps. Google's guidance is plain about access: "To edit your business info, you need to sign in to the Google Account that's linked to your Business Profile," after which you edit directly in Search or Maps. It is equally plain that edits are not instant, because "We review your changes before we update them live on your profile." If the profile is not yours to sign in to, the same help pages point to two routes: suggest an edit on the profile, or claim it.
The part almost nobody knows sits one page deeper. Google states that it "gathers info from different sources, like user reports and licensed content," and that if those sources report your information as incorrect or outdated, "Google may update your profile." Inside the profile editor, blue text marks what Google changed rather than you. Google also says you cannot manage every one of those updates from the profile. Read that as an operating instruction: a Business Profile is not a form you submit once. It is a record that other people write to.
Two. The knowledge panel. The panel about your company or about you personally is a separate object with a separate process. Google's help documentation describes claiming it by finding the panel, choosing "Claim this knowledge panel," and verifying yourself through an official account such as YouTube, Search Console, X or Facebook. Once verified, you use Feedback below the panel, flag the specific fact, describe what it should say, and, in Google's words, "Provide links to web pages that can be used to verify this information." That last line is the whole discipline in miniature. You do not assert the correction. You evidence it.
Three. Pages that are wrong and not yours. For the abandoned bio on a former partner's site, or a directory entry describing a service you retired, Google publishes the Refresh Outdated Content tool. It is for pages you do not own, and only where the page or image "no longer exists" or is "significantly different" from what the search result shows. It is not a recrawl request, and it is not for your own site, where verified owners use the Removals tool instead. Processing takes a few days, and an approved request refreshes what Google shows rather than rewriting the page. The wrong page still exists. If the stale claim is still live on it, you are back to asking its owner.
All three correct records that Google's own products draw on, including Search and Maps. What they do not do is travel. The Google and OpenAI documentation cited here describes no route from your Business Profile into ChatGPT, which is the argument we set out in full on whether your Google Business Profile helps you in ChatGPT. Google is also not the only listing record. Microsoft and Apple keep their own for their own map and search products, and each takes the same treatment: claim it, correct it, check it again later.
What ChatGPT Can Actually Reach
ChatGPT reads a different pile, so it takes different filings. Three areas are worth reviewing, and the first one is the one that quietly breaks.
Crawler access. OpenAI's crawler documentation names OAI-SearchBot as the crawler "used to surface websites in search results in ChatGPT's search features," and states the consequence bluntly: "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers." Its recommendation is to allow OAI-SearchBot in robots.txt. Blocks are rarely deliberate. They arrive as a security rule, a firewall setting, a bot filter switched on by whoever manages the site. The page reads perfectly to you while staying unavailable for ChatGPT search to crawl and quote. Check it before you rewrite anything, because a corrected page that crawler cannot fetch is a corrected page it cannot quote.
Worth separating from that: GPTBot is a different crawler with a different job, described as making foundation models more useful and safe. OpenAI's documentation is explicit that ChatGPT-User "is not used to determine whether content may appear in Search." Blocking training and blocking search are two decisions, and firms regularly make the second one by accident while intending the first.
Licensed partners. On July 23, 2026, OpenAI and Yelp announced a deal licensing Yelp's reviews, ratings, photos and business information into ChatGPT's local answers, with Yelp branding and links shown alongside. For a professional firm that has ignored Yelp for a decade, that changed the status of the page overnight. It is no longer only a directory listing. It is a correction surface with a documented route into one of the assistants your buyers use, so where ChatGPT draws on Yelp for a local answer, an outdated Yelp page can shape what a stranger reads.
Your own pages. Put the plain facts in one obvious, crawlable place: legal name, trading name, city, what you do now, who you serve, who works there, and what you stopped doing. Then make the structured data say the same thing, so your identity is stated rather than inferred, which is the work we publish as Entity Identity. Consistency will not force a result. It removes the conflicting signals that let a machine pick the wrong version of you.
Why One Fix Lands on Five Clocks
Here is the asymmetry that makes this work feel unfair. The correction is one afternoon. The arrival is somebody else's calendar, five times over.
On August 11, 2026 we ran our own instrument: ten buyer questions across five engines, fifty answers, every answer saved word for word. On one question, "Who can audit what AI assistants say about my business?", the five answers barely overlapped. Perplexity's sonar model named one set of vendors, Gemini 2.5 Flash another, GPT-5.2 another, Claude Sonnet 4.6 another, Grok 4.3 another, all with live web access on, all on the same day. Only two companies were named by more than one engine, and neither by more than two. That is one question on one day, so hold it as an observation rather than a law. The observation still travels. Five engines, five separate bodies of evidence, and no shared inbox where one correction lands for all of them.
Two practical consequences follow. Check per engine rather than averaging, because an average hides the engine that still has you wrong. And expect uneven timing. An assistant that searches while it answers can reflect a corrected page far sooner than one answering from memory, and none of the providers discussed here publishes a date by which your fix will appear. Google at least documents ranges, from days at the fast end to months at the slow one, which is why we wrote out how long AI visibility work actually takes using its own published wording.
One more honest boundary, because it decides what you do after the filing. Being accurate is not the same as being named. In a vendor study by Ranqo covering more than 100,000 AI answers about 100 or more brands between March and May 2026, household-name brands appeared in 73 percent of relevant unbranded answers, mid-market brands in 44 percent, and small or niche brands in 11 percent. One vendor, one sample, and it measures appearance rates rather than the effect of any repair. The shape is still the useful part. Correcting the record clears the reasons to skip you. It does not by itself buy a place in a crowded answer. That is why the plan after a cleanup targets the questions where the big names are absent and your experience is deeper.
So date everything. Save the wrong answer verbatim with the engine, the date and whether search was on, make the fix, then re-ask the identical questions later against the same list. The method is the one we published in how to measure whether AI visibility work is working: an appearance rate with a denominator, never a screenshot.
When a Correction Needs a Lawyer
Nearly all of this is administrative. A few cases are not, and they are worth naming before you publish a rebuttal in the heat of the afternoon. Nothing here is legal advice, and the rules differ by country, state and profession, which is the point.
Get counsel involved when the answer alleges misconduct, wrongdoing or a legal outcome that did not happen, rather than simply having a fact out of date. Get counsel involved when the correction would touch a confidential matter, a sealed proceeding or a client you cannot name, because the accurate version can breach something the wrong version did not. Get counsel involved when your profession's advertising rules govern what you may claim about results, which is routine in law, medicine, financial advice and insurance, where the fix has to survive a regulator as well as a search engine. And get advice when the dispute is about a name, a former partner or a business that ended badly, because those corrections tend to be contested by a person rather than processed by a queue.
In any of those cases, preserve before you act: the exact question you asked, the answer word for word, the engine and model, the date, and whether web search was on. A screenshot alongside the text costs nothing. Then take advice, then correct in public.
The reason to be careful is simple. A public correction can be indexed, copied and repeated, so publish only what you are content to have durable. A rebuttal page written while you are angry becomes evidence with your name on it, and it is the version the engines can find long after you have calmed down. Keep the record dated, factual and boring. Boring survives.
Everything else is a filing job with an order to it. Start with the surface your buyers and the machines consult most, which for anything involving a location is usually your Business Profile. Make your own site state the current facts plainly. Retire or correct the abandoned profiles that still describe a company you no longer run. Confirm the assistants can actually fetch your pages. Then wait, and re-ask.
If you want the full reading before you start, the free Recommendation Check asks your buyers' questions across five engines and scores your record against 109 checkpoints in eight zones. Scores run 0 to 109. Ninety and up means AI treats you as a primary source. Under 35 means you are effectively invisible. There is no discovery call attached to it, and if the report says you are in good shape, you keep it.
Want to Learn More?
Probably Genius builds and operates the recommendation layer around expert-led businesses, then measures it in the open. That is the job inside The Answer: twelve done-for-you thought-leadership articles a month, planned to create approximately sixty clear Knowledge Entries, your expertise structured so machines can verify it, corroboration built beyond your own pages, and a monthly probe across five engines with every answer saved word for word. The published methodologies show how the work runs, including the 109-Point Diagnostic behind the free check. We never promise what a model will say. We make you easier to verify, then we test it honestly.
You're probably a genius at what you do. We make sure AI gets the memo.