Wrong information about your business in AI: the five step correction

When an assistant states something wrong about your business, correct the source rather than arguing with the output. An AI answer is assembled from pages, listings and profiles that exist somewhere public, so almost every wrong fact traces back to a page you can find and change. The sequence that works is: capture the error with evidence, classify it, find every source carrying it, correct them at origin, publish the right version in plain dated text on your own site, then re-test and report to the platform if it persists. This page takes you through each step.

Step 1: capture it properly

Before you fix anything, build a record you can act on and show to somebody else.

• Note the exact prompt, the assistant, the date and time, and whether you were signed in.
• Copy the answer verbatim, not a paraphrase.
• Screenshot it, including the citation list if one is shown.
• Ask the assistant a follow up: did you search the web for that, and which pages did you read. Every major assistant will answer, and that single question usually hands you the source.
• Run the same prompt twice more in a temporary chat with memory off. An error that appears once in three runs is a different problem from one that appears three times out of three.

Keep these rows in one sheet. Over a quarter it becomes the evidence log that shows a correction worked.

Step 2: classify the error

There are five kinds, and each has a different fix. Working out which one you have saves a week.

Kind	What it looks like	Where the fix is
Stale fact	An old price, an old address, a product you retired, a director who left	Every public page still carrying it, including feeds and affiliate pages
Wrong status	Says you are closed, merged or no longer trading	Business profile, map data, directory listings
Confused entity	Your facts mixed with a similarly named company	Disambiguation: distinct naming, category and location in every listing
Missing fact filled in	An accreditation, a founding year or a team size that nobody published, so the model inferred one	Publish the real figure plainly, with a date
Someone else's claim	A review, forum thread or article stating something incorrect	The platform's own correction route, plus your own dated statement

The fourth row is the one people miss. If a fact about your business exists nowhere public, an answer will still be produced, built from whatever is nearest. Publishing the number is the fix.

Step 3: find every source carrying it

One corrected page is rarely enough, because the wrong version usually lives in several places.

• Search the wrong fact itself in quotation marks, with and without your business name.
• Search your business name and open everything on the first three pages of results that is not yours.
• Check the five sources most often cited in your own prompt rounds first, since those are demonstrably being read.
• Check your own site for the old version: an archived blog post, a PDF, a footer, a careers page, a press release.
• Check your feeds and any marketplace or reseller listings, because a stale feed outlives a corrected web page.

List them all with an owner and a contact route before you start sending emails. Our 30 query method for building this list is in the fact accuracy audit with evidence log.

Step 4: correct at origin, then state it on your own site

Work outward from the sources with the most authority, and do both halves.

• Your own pages first. Correct the fact everywhere it appears, including old posts, and add a review date.
• Your structured data. Update Organization or LocalBusiness markup so it matches the visible text, which is what Google asks for.
• Profiles you control. Business profile, review platforms, directories, trade bodies, your social pages. These are structured and easy for a model to read, which is why they carry weight out of proportion to their traffic.
• Pages somebody else owns. A short, factual email: here is the fact, here is the page on our site stating it, here is the date. No argument, no complaint.
• Then publish one plain page stating the correct version. Title it as the question somebody would ask, answer it in the first two sentences, give the number and the date, and keep it as the canonical source you point everyone at.

That last page does double duty. It is the reference you send to anybody correcting a listing, and it is the quotable passage an assistant can lift next time. The structure is in how to write a page an assistant will quote.

Step 5: re-test, and report if it persists

Give it two to four weeks for a fetched answer, since the engines need to re-crawl the corrected pages. Then run the same prompt three times in a clean session and compare against your log. Answers written from model training move on a longer timescale, which is why the public record is the thing you change.

Where a wrong statement persists after the sources are corrected, every major assistant has a feedback route in the answer itself, usually a thumbs down with a comment box, and Google's AI surfaces have a feedback link on the overview. Use it with the specifics: the prompt, the wrong claim, the correct claim and the URL that proves it. The documented route for Google is in how to document and report an incorrect AI Overview.

The two special cases

Confused with a similar name. This needs disambiguation rather than correction: a distinct legal name, a specific category, a stated location and a consistent description everywhere, so the two businesses stop looking like one. The audit is in the disambiguation audit.

Said to be closed. Treat this as urgent, because it costs you enquiries every day it stands. Check status fields on your business profile, your map listing and every directory, in that order. The playbook is in when AI says your business is closed.

This page is the correction process for a wrong fact. For an assistant that has never heard of you at all, which is a different problem, see what to do when AI has never heard of your business. For the ChatGPT specific route, see what to do when ChatGPT gets a fact about you wrong.

wrong information about your business in AI, in practice: the usual questions

Why does AI say wrong things about my business?

Usually because the wrong version is published somewhere public and the right version is not, or because the fact was never published at all and the model filled the gap. Both are addressed by putting the correct fact in plain text with a date and getting the stale copies corrected.

Can I ask an assistant to remember the correction?

A correction inside your own chat applies to your session, not to what other people are told. The durable fix is the public record, which is why the work is on the sources rather than in the conversation.

How long does a correction take to show up?

Answers that come from a live fetch can change within two to four weeks of the corrected pages being re-crawled. Answers written from model training move on the timescale of the next model, so keep the public record right and re-test on a schedule.

What if the wrong fact is on a site that will not change it?

Outweigh it. Make sure your own page, your structured data and three or four independent profiles all carry the correct version with a date. A single stale page against five current ones is a much weaker source than it was.

Should I use a legal letter?

Start with the short factual email, because most editors and directory operators correct a verifiable fact without any pressure. Keep stronger steps for a claim that is both damaging and refused, and let your evidence log do the work.

What to do first

• Run your five buyer questions plus a plain question about your company in a temporary chat, and log every wrong statement verbatim with a screenshot.
• Ask each assistant which pages it read, and write those URLs beside the error.
• Classify each error using the table above so you know which fix it needs.
• Correct your own pages, your structured data and the profiles you control this week.
• Publish one dated page stating the correct facts, send it to every third party carrying the old version, and diarise a re-test in three weeks.