When the record is wrong
An engine does not keep a record of you that someone can edit. It assembles a description at the moment of the question, out of sources — and the ones that decide the answer are usually not on your domain.
The short answer
There is no correction form, no support ticket and no verified-business panel. That is the bad news. The good news is that the thing you can change is concrete, it is a finite list, and most companies have never looked at it.
First, work out which kind it is
“The AI said something wrong about us” is four different problems wearing one sentence. We split them because the repair is different for each, and because the last two vanish if you normalise the text before comparing — which is how most tools miss them.
| Kind | What it looks like | What fixes it |
|---|---|---|
| Contradicted | A source states something that is simply not true now, and your own pages say otherwise. | A correction request to that source. Slow, human, and the highest-value thing on this page. |
| Outdated | It was true. It is being read as current because nothing marks it as old. | Publish the current version with a date, somewhere already trusted, so the fresh record outweighs the stale one. |
| Spelled differently | The same thing written two ways — capitalisation, spacing, a legal suffix that comes and goes. | Pick one spelling and push it everywhere. An engine is less likely to treat the two as one entity than you assume. |
| Fragmented | Several legitimate names for the same product or programme, all in circulation. | Choose the primary name and make every other record point at it, so authority accumulates on one of them instead of splitting. |
The last two are the ones nobody reports, because any system that lowercases text before comparing makes them disappear. We found that the hard way and had to compare twice — once normalised to find the match, once on the original to see the difference.
Then fix them in this order
The instinct is to edit your own website, because that is the part you can change this afternoon. It is usually the least valuable move available.
Ask the engines the questions your buyers ask and record what comes back with the citations attached. You are not looking for your mention rate yet. You are looking for the list of domains the answer was built from, because that list is your worklist.
A wrong fact on a high-authority source that gets cited constantly matters more than three wrong facts on pages nothing reads. Score per source and take the worst one, rather than averaging — an average lets one well-behaved source hide a bad one.
Registries, databases, directories, trade press. A form, an email, a wait. This is the slow part and the part that moves the answer, because corroboration between independent sources is what the engine is weighing.
Last, not first. One wording, one spelling, one primary name — and in the page text rather than inside an image or a graphic. A diagram of your product range is invisible to the thing you are trying to persuade.
What this looks like written down
This is the shape of what we hand over: each disagreement, which sources carry it, what the approved version is, and what to do about it — ranked, so there is an obvious first thing.
severity 0.21 mechanism
should say: selective Class I inhibitor
differs : “oral 1/2/3 inhibitor” trade press
differs : “pan-inhibitor” database
severity 0.18 pipeline
should say: lead asset active; second asset discontinued
stale : “second asset in development” database
severity 0.09 company_name
spelled : two capitalisations in circulation
alias : with and without the legal suffix
Every line links to the page it was read from, with the date.
Values above are from our own test fixture, not a customer. We do not publish anyone’s diagnostic, and nothing about a company we have measured appears anywhere on this site.
Where this stops
When an engine answers from training rather than from a live search, there is no citation and no source to correct. That changes when the model is retrained. Anyone selling you a fix for this is selling you something that does not exist.
If the record is accurate and unflattering, this is not the service you want. We correct what is wrong and we do not launder what is right — no planted sources, no invented corroboration, no reviews from people who do not exist.
Questions
There is no correction form and no support queue for this. The engines do not hold an editable record of you — they assemble a description at the moment of the question from whatever they trained on and whatever they just read. The only editable thing is the sources, which is why the work is almost entirely off your own site.
Days for anything an engine re-reads on a live search, which is most of what you can influence. Weeks for a registry or database correction, because a person on their side processes it. Not at all for what the model recalls without searching — that changes when the model is retrained, which you do not control.
That is the normal case and it is the part worth paying for. Most corrections are a form, an email and a wait, and they outrank anything you publish on your own domain because corroboration is what an engine weighs. There is no API for it, which is why most vendors skip it.
Usually not, and sometimes it makes things worse. A deleted page can stay in a third-party copy, in the training data, and in whoever quoted it. The reliable move is to publish the correct version somewhere an engine already trusts, so the record disagrees with the stale version rather than falling silent.
Related reading
The other half: once the record is right, what makes a page the one an engine quotes.
How to tell whether the correction worked, and why one check proves almost nothing.
The definition, and the five things that move whether your name is in the paragraph.
The holdout method, which separates your correction from a market that moved anyway.
Start here
The free diagnostic: how public sources describe your company and your main product today, where they disagree, and which disagreement to fix first. Every line links to the page we read it from. Nothing is published anywhere without your written approval.