01The Five Types of Error
Five different errors, five different fixes
The Short Version
Name the error before trying to fix it. Outdated data, wrong attribution, mistaken identity, pure invention and merged-with-a-competitor each have a different origin and a different deadline.
Outdated data is the easiest: the address, the phone, the partner, the capacity in tonnes per month — all correct at some point and stale now. Wrong attribution is when the model says you make something you have never made, usually because your site lists a line you discontinued or a supplier partnership someone described loosely. Mistaken identity happens with common names: two companies called Metalúrgica Progresso, in different states, merged into one answer. Pure invention is the model filling a gap with a plausible sentence — a certification you do not hold, a founding year nobody published. And merging with a competitor happens when your names appear together often enough in comparison lists that the model attributes one company's attributes to the other.
The practical consequence of naming the type: outdated data and wrong attribution are fixed at the source and disappear fast. Mistaken identity needs disambiguation — publishing what separates you from the homonym, explicitly. Pure invention is the only one where the real answer is filling the gap: the model invented because nothing authoritative existed to read, so publish the fact and the invention loses its slot.
A model does not lie on purpose and it will not argue with you. It repeats the most available version of your company. Change what is most available and the answer changes with it.
02Where the AI Got That
Find the source before touching anything
Ask the assistant directly: "where did you get that information about this company?". Engines with live search return the links. That list is your work order — in Brazilian manufacturing the culprit is almost always one of these five:
Source 01
Your Own Site, Never Updated
The "About Us" page still says forty employees and one plant. The footer carries the old address. The catalogue page still lists the line you dropped in 2021. The model is reading you correctly — you are the one publishing the wrong fact. Check the footer, the About page and the catalogue before blaming anyone else.
Most Common
Source 02
Directories and Marketplaces From Ten Years Ago
Supplier listings, B2B marketplaces and association pages you registered on once and forgot. They rank well, they never expire, and they still show the old phone number. Search your company name in an anonymous window and go through the first three pages of results — the stale listings are right there.
Third Party
Source 03
An Old News Item That Never Aged Out
A 2016 trade-press piece about the acquisition that did not go through, or the labour lawsuit, or the plant that later moved. News pages carry high credibility for a model and rarely get corrected. The counterweight is publishing current information that is more detailed and more recent than the article.
Press
Source 04
An Unclaimed Business Profile
Business profiles, map listings and review pages created automatically or by a former employee. Anyone can suggest an edit to an unclaimed profile, and those edits propagate. Claiming it costs one verification and closes an open door.
Profile
Source 05
A PDF From 2014 Still Indexed
The old price list, the old spec sheet, the old capacity brochure — sitting in a folder on your own server, undeleted, fully indexed. It contradicts your current pages and the model cannot tell which one is newer. Remove or replace the file and request reindexing.
Legacy
03Correction Protocol
The correction protocol, in order
Order matters. Sending feedback to the assistant before fixing the source achieves nothing — the next re-read restores the error. Work from the source outwards:
Five Steps, From Source to Assistant
Step 1: Fix the Primary Source
Your site first — footer, About page, catalogue, and any legacy PDF that contradicts the current version. Delete the stale files instead of leaving them orphaned, and request reindexing in Search Console for the corrected pages.
→
Step 2: Publish the Canonical Facts Page
One page carrying legal name and trade name, founding year, plant addresses, product lines currently in production, discontinued lines named as discontinued, certifications with their numbers, and the last update date. This is the address you point everything at.
→
Step 3: Claim and Correct the Third Parties
Business profile, directories, marketplaces, association listings. Claim what can be claimed; e-mail the editor for what cannot, attaching the link to your canonical page. Half of these are handled in an afternoon by one person.
Steps 4 and 5:
Report the wrong answer inside each assistant using its own feedback control — the thumbs-down or report link under the response — describing the error and pasting the link to your corrected page. Then wait and re-test on a schedule, not daily: engines with live search usually stop repeating the error between two and thirty days after the source is reindexed, while an error baked into training memory only leaves in the next training cycle, which takes months and does not depend on you. In the meantime, the live-search route already gives the buyer the corrected version, which is what actually matters commercially.
One warning about step 2: a facts page only works if it is readable as text and the crawlers are allowed in. If your robots.txt blocks the AI crawlers, you have just published a correction nobody can read — the reasons and the file to check are covered in why ChatGPT does not recommend your company.
04How to Monitor Afterwards
Correcting once is not enough
The error came back into the answer because something on the web still says it. New stale sources appear every year: a directory that copies your old data, a news piece, a former employee's profile. Treat this as a quarterly routine, not a one-off project. Twelve fixed questions, written once, re-run every three months in a clean session, with the answer pasted into a spreadsheet alongside the date. It takes half an hour per quarter and finds the wrong fact before a customer does.
A wrong fact about your company circulating in AI answers costs more than absence. Absence means the buyer never heard of you; a wrong fact means they heard something and decided against you.
The set of questions to use, how to word them so the answers stay comparable, and how to log the result are in how to find out what ChatGPT says about your company. A specific and very common case of mistaken attribution — the AI naming your distributor instead of your plant — is handled in the distributor shows up in AI and the manufacturer does not. For a continuous measurement instead of a manual sheet, see the share of model audit. The whole pain sits in my company does not show up on Google or in AI.
05FAQ
FAQ
How do I fix wrong information an AI says about my company? +
Correct the source, not the answer. Ask the assistant where it got the information, fix the page that carries the wrong fact, publish a canonical facts page on your own domain, claim the outdated profiles and directories, and then report the answer through each assistant's feedback control with the link to your corrected page. Engines with live search usually update between two and thirty days after reindexing.
Can I sue an AI company over wrong information about my business? +
It is a question for your lawyer, and it is rarely the fastest route. Every major assistant offers a reporting channel and a correction request form, which resolves in days to weeks; a legal dispute takes far longer and does not by itself remove the wrong source from the web. Report it and correct the source in parallel, and keep dated screenshots of the wrong answer if the damage is material.
How long until the AI stops repeating the error? +
It depends on where the error lives. If it comes from a page the assistant reads live, expect two to thirty days after the source is corrected and reindexed. If it is stored in training memory, it only leaves in the next training cycle — months, and outside your control. In both cases correcting the source is the only lever you actually hold.
The AI says my company closed down. What do I do? +
Usually there is an old news item, a bankruptcy filing for a company with a similar name, or a business profile marked as permanently closed. Check the profile first — that flag propagates fast and is reversed with a claim and a verification. Then publish a dated facts page showing current operation and report the answer with that link attached.
The AI confuses my company with another of a similar name. How do I separate them? +
Disambiguate explicitly instead of just repeating your name. Publish pages that always carry name plus city plus state plus segment together, use the legal name and the tax registration number on the facts page, and write one sentence stating you are not connected to the homonym company. Repetition of the distinguishing pair is what teaches the model the separation.
The AI quotes an old price for my product. How do I remove it? +
Find and delete the indexed source — usually an old price list in PDF on your own server, a marketplace listing or a distributor page. Deleting the file is not enough if the URL still returns content, so remove it properly and request removal in Search Console. As a rule, do not publish industrial prices as static text you will not maintain; publish the pricing criteria and route the number through a quote.
Do I have to notify each AI separately? +
For the report, yes — each assistant has its own feedback channel and they do not share it. For the actual correction, no: fixing the source and publishing the canonical page serves all of them at once, because they all read the same web. Do the source work once, then spend ten minutes reporting in each assistant that showed the error.
How do I stop the AI from inventing things about my company? +
Invention fills gaps. The more of your basic facts are published clearly and consistently — lines in production, capacity, certifications with numbers, plants, founding year, all on one dated page and repeated across independent listings — the less room there is for the model to guess. You cannot switch invention off, but you can leave it nothing to invent about.
Want to know what the AIs are getting wrong about you?
We run the fact questions across the main assistants, trace each wrong answer back to the source that feeds it, and hand you the correction list with the responsible party and the expected deadline for each item. Bring your domain and the list of profiles you remember creating.
