How Do You Correct False Information About a Company?

You correct false information about a company at its sources, not inside the AI system — there is no edit button in ChatGPT, no correction form that rewrites what an AI answer says. AI systems assemble their descriptions of a business from published information: the company’s website, directory listings, review platforms, news coverage, knowledge panels, and the copies of all of those that other sites have made. The correction discipline is therefore a publishing discipline: locate every source carrying the falsehood, correct it at the origin first, align everything that corroborates it, and then verify, on a schedule, that the answers have actually changed.

That sounds slower than business owners want it to be. It is also the only version that works, and understanding why it works this way is the difference between a correction campaign that converges and one that chases its tail.

Step 1: Find every source carrying the falsehood

A wrong fact about a company almost never lives in one place. It lives in a family tree. The original error — an outdated listing, a misreported news item, a directory entry scraped from somewhere else, a former employee’s old profile page — gets copied by aggregators, echoed by data brokers, quoted by other directories, and absorbed into the indexes that AI systems read. Correcting the page where you first noticed the error while six copies of it remain live is the most common reason corrections “don’t take.”

So the first step is an inventory, not an edit. Document what AI systems currently say about the company, in their own words, with dates. Then trace each false claim to the sources that could be feeding it: the company’s own site (including old pages nobody remembers publishing), its claimed and unclaimed listings, review and map profiles, industry directories, press coverage, and structured sources such as knowledge panels and Wikidata-style records. This is detective work, and it is exactly the work of a Business Entity and Source Verification engagement: a discrepancy register that lists each conflicting claim, each source carrying it, and which version is authoritative.

Step 2: Correct at the origin first

Copies of copies are the hardest to fix and the least important to fix first. Origins are the opposite. If the company’s own website still states the old address, the old ownership, or a service it stopped offering in 2019, every downstream correction is built on sand — aggregators refresh from primary sources, and the error will regrow.

The origin tier usually includes: the company’s website (all of it, including archived service pages and PDF menus or brochures, which AI systems read as happily as HTML); its primary business profiles; and any page the company controls that other sources visibly copy. Correct these completely and consistently — the same wording for the same facts everywhere, because AI systems treat agreement across sources as a signal and disagreement as a coin flip.

Step 3: Align the corroborating sources

Next tier out: the directories, map listings, review platforms, and industry databases that repeat the company’s facts. Claim the unclaimed listings — an unclaimed profile is a source anyone’s data can overwrite. Submit corrections through each platform’s own channel, and expect each platform to run on its own timeline, in its own format, with its own idea of proof.

Where platforms offer a formal correction route, use it and keep the receipts: knowledge panel feedback mechanisms, directory claim-and-verify flows, map listing edits. AI Reputation Assessment work typically surfaces these identity-level problems — a business conflated with a same-named competitor, a cloned listing, a profile describing a company that closed a decade ago — and each has a different correction path. There is no universal one.

Step 4: Document before and after, with dates

A correction you cannot demonstrate is a rumor. Record the state of the record before you change anything: screenshots and transcripts of what AI systems answer today, the source inventory, the dates each correction was submitted and accepted. Then re-test on a schedule — not daily, because propagation is slow and answer variability will mislead you, but at defined intervals, with the same questions, recorded the same way.

This documentation does three jobs. It tells you whether the correction is converging. It gives you evidence if a platform, a data broker, or — in serious cases — a regulator or a court needs to see the timeline. And it disciplines the process: teams that record before-and-after stop declaring victory after editing one website page.

The hard truths

Corrections propagate unevenly. Search indexes refresh on their own schedules. AI systems that answer partly from training data may carry an old fact long after every live source is fixed — a model cannot cite a correction it has not ingested. Some answers will improve in weeks; some traces of an error can persist for far longer, and nobody honest will promise you a date.

Some copies are unreachable. Defunct aggregators, scraped-content farms, archived pages, screenshots in other people’s articles: a long tail of copies exists that no correction request will ever touch. The practical goal is not erasure. It is making the correct, consistent record so dominant across the sources that matter that the unreachable copies stop being the version AI systems assemble.

The corrected record is the win, not a takedown. Owners often ask how to “remove” a falsehood from AI. Removal implies a delete key that does not exist. What exists is replacement: a true record, published in retrievable form, agreed across independent sources, maintained over time. AI answers follow the weight of the available record. Change the weight and, unevenly but reliably, the answers follow.

What this looks like in practice

A regional dental practice discovers that AI answers describe it as offering orthodontics — a service it dropped years ago — and list a partner who retired. The inventory finds the claim on the practice’s own forgotten services page, two directory profiles built from that page, a health-listing aggregator, and a “meet the team” article on a local news site from 2018. The practice corrects its own page the same day, claims and fixes both directories within the week, and submits corrections to the aggregator with documentation. The news article is not corrected — the outlet does not update archive pieces — but it stops being decisive once every current source agrees against it. Re-testing at 30, 60, and 90 days shows the orthodontics claim fading unevenly across systems, then disappearing from most, with the dated record showing exactly when each answer changed. That uneven fade is what a successful correction actually looks like.

Misconceptions

“There’s a form to fix what ChatGPT says about us.” There is not. Platforms offer correction channels for their own listings, panels, and map data — and those matter — but the AI answer itself is assembled from sources. Fix the sources.

“We corrected our website, so we’re done.” Your website is one source among many, and often not the one being copied. If the directories still disagree with it, you have published an argument, not a correction.

“A lawyer’s letter will remove it.” Legal process has real uses — defamation, impersonation, and trademark confusion among them — but a demand letter does not edit an index, and most false business facts are stale data, not libel. The correction discipline resolves the data problem; legal review is a parallel track for the cases that genuinely involve one.

“Once it’s fixed, it stays fixed.” Listings get overwritten by data feeds, profiles get re-scraped, businesses change. Correction is a maintenance state, not an event — which is why monitoring exists as a discipline of its own.