When AI gets your company wrong: three causes and what actually fixes them

Why AI invents things about you

Language models like ChatGPT, Gemini or Perplexity don't hold facts the way a database holds a row. They predict the next token from patterns in what they were trained on plus whatever they retrieved for this particular question. When there is no solid public record of your company, the honest output would be an admission of ignorance. What comes out instead is the answer that would be typical for a business like yours, phrased with exactly the confidence of a verified fact. The industry word is hallucination, and the confidence is the dangerous half: in the Tow Center's test, ChatGPT misidentified 134 of 200 articles while flagging any uncertainty just 15 times, and never once declined to answer.

Then there is the clock. Profound looked at roughly 730,000 US English ChatGPT conversations from late 2025 and found only about 18 percent triggered any web search at all; the rest were answered from training data alone. So a move, a new managing director, a service you dropped, a change of ownership: none of it exists for the model until something forces it to look. The answer is accurate about the company you used to be. A firm that merged, a trades business under new ownership, a clinic that opened a department all get this treatment, and they get it on every query until the public record catches up.

The third cause is your own sources disagreeing. When three phone numbers, two addresses and two spellings of your legal name are all in circulation, the model has to pick one. It tends toward the version that appears most often or reads most authoritatively, which is not the same as the current one. These are the errors that embarrass you worst, because nothing in the answer looks wrong. It is fluent, specific, internally consistent, and out of date.

The errors that actually show up

Not every error costs the same. Mistaken identity is the nastiest, because the answer holds together. An engineering firm shares a name with an online shop, and the model quietly merges the two into a company that doesn't exist: your services, their reviews, their returns policy, one plausible paragraph. Nobody reading it has any reason to doubt it, which is why you will only ever catch it by asking the systems yourself.

Then the ordinary factual errors: a price you changed two years ago, a certification that lapsed, an award you never won, a service you don't offer. The reverse happens just as often and is far easier to miss. The model leaves out the thing you actually do best, because nowhere on the open web is it stated plainly on a page a machine can parse. For a service business that is the expensive version. You are not wrong in the answer, you are absent from it, and the buyer never puts you on the shortlist.

The heavy cases touch your reputation. A model that generalises one resolved complaint into a claim about your quality, or that attaches an insolvency rumour to the wrong company with a similar name, does real commercial damage. Those are rarer than the mundane errors, and they are the ones to document the moment you see them, because the answer in front of you may not be reproducible tomorrow.

How to find out what the systems say

The first step is dull and almost nobody does it: ask the systems yourself. Put the questions a buyer would actually type into ChatGPT, Gemini, Perplexity and Copilot. Who are you, what does this cost, where do you work, are you certified. Save the answers verbatim rather than summarising them, and note the model and the date. Then ask again in slightly different words, because phrasing changes what gets retrieved, and two wordings of the same question can describe two different companies.

Chance sampling tells you nothing. Fix a list of questions, run it on the same day each month, and keep the answers side by side. That is how you learn whether an error is new, persistent or already fixed, and whether the work you did at the source moved anything at all. Without the log you are guessing, and you find out about a bad answer when a prospect quotes it back to you on a call.

Read the citations wherever you get them. Perplexity and Copilot usually show what they pulled from, and Google's AI Overviews link out too. When an abandoned directory listing or somebody else's profile page turns up there, you have found the actual source of the error instead of theorising about it. Ahrefs, looking at 1.4 million real ChatGPT prompts, found the model cites only about half the URLs it retrieves, so treat the visible list as a useful sample of what it read, not the whole of it.

Fix the cause, not the sentence

The most effective fix is unglamorous: say the same thing everywhere a machine can read it. One legal name, one trading name, one address, one phone number, one service list, identical on your site, in the directories, in map listings and on your social profiles. Every variant is another candidate answer, and a model with two candidates will sometimes hand out the wrong one. This is hygiene rather than strategy, and it settles more cases than anything clever you do afterwards.

Structured data helps, as long as you are honest about what it does. Schema.org markup on your contact, service and about pages makes the facts extractable, which kills the class of errors where a model guesses at price, hours or location. Google has said since 2018 that schema is not a direct ranking factor, and its AI-features guidance is explicit that no special markup, schema or AI-specific file is needed to appear in AI Overviews or AI Mode. Treat it as making yourself easy to quote correctly, not as a lever on visibility.

Then go after the sources that carry weight. In Profound's sample of ChatGPT conversations, Wikipedia was the most-cited domain at about 5 percent of all citations, with Reddit next at around 3 percent. Official registers, the large trade portals and established review platforms behave the same way. One wrong entry there outweighs ten correct pages on small sites, so fix in order of reach. The evidence points the same direction: across roughly 75,000 brands, Ahrefs found how often a brand is mentioned around the web correlates with AI citation at about 0.66, roughly three times as strongly as backlinks do.

When the claim is damaging

When a model states something both false and harmful, capture it before you do anything else: a full screenshot, the exact prompt, the model and version, the date and time. Answers are not stable, and the one you just read may be unreproducible within the hour, which is precisely the moment your complaint stops being actionable. Recording it takes a minute, and it is the step people skip.

Every major provider has a route for reporting a bad output; OpenAI, Google, Microsoft and Perplexity all accept feedback on specific answers. Use it plainly: here is what the answer said, here is what is true, here is the page that proves it. Do that and fix the underlying source on the web in parallel, because a report on its own rarely changes what the system says next month. The report addresses one answer. The source addresses the reason.

If a false, damaging claim persists and you can point to lost business, get a lawyer involved. Defamation and unfair-competition law reach automated statements in most jurisdictions, though who is liable for a generated sentence is still being settled in the courts and the answer differs by country. This is the last step, not the first. It is also only available to you if you kept the documentation.

What a correction realistically takes

Be clear-eyed about the mechanism: there is no record to edit. You cannot open the model and change the row about your company. You change the information around it and wait for the systems to notice. Between fixing the source and seeing the answer move, expect weeks, sometimes considerably longer, depending on how and when a given system refreshes what it knows.

The two kinds of system move at different speeds, so track them apart. One that retrieves live pages can pick up a change within days. One answering from training data waits for the next update, and per Profound's sample the large majority of ChatGPT conversations never trigger a search at all, so training data is doing the talking more often than you would guess. Ahrefs also found the median cited page is around 500 days old, which cuts both ways: slow to reflect your fix, and what you publish today will still be answering for you long after you have forgotten writing it.

Don't aim for total control; you won't get it. Aim for a public record clean and consistent enough that the most probable answer happens to be the true one. Get that right and wrong answers become rare and shallow instead of routine. They will not hit zero, and any vendor promising otherwise is selling something they cannot deliver.

The order to work in

Don't start with a tool, start with an inventory. Write your correct core facts down in one place: legal name, trading names, address, phone, service list, service area, certifications, the people who speak for you. Then check what the main systems say and where on the web they could plausibly be getting it. Until you know where the public record diverges from the truth, you are optimising in the dark.

After that, work in waves. Your own site and the big directories first, then the high-reach third-party sources, then the long tail. Re-measure after each wave so you can tell which change did the work. That loop, measure and fix and measure again, is the job itself. It is not a project with an end date, because the sources keep moving whether or not you are watching them.

Then make it routine. Once a month, the same questions, compared against last month's answers, is enough for most small and mid-sized companies. Monthly is often enough to catch an error while it is still one wrong sentence, and cheap enough that you will actually keep doing it past the second month.

• Write your core facts down in one place
• Ask four or five systems the questions a buyer asks
• Make your site and the big directories agree
• Fix the highest-reach third-party sources first
• Re-run the same questions monthly and compare

Where a wrong answer costs most

The risk is not spread evenly. In health, finance and law a wrong answer can do damage the same day: when a language model credits your practice with a treatment you don't provide, or gives your firm a specialism it never had, you are looking at a compliance conversation and a lost referral rather than a typo. These are the sectors where close monitoring pays for itself, because one wrong sentence can outweigh a year of accurate ones.

In retail and hospitality the damage is operational: yesterday's opening hours, last year's prices, a location you closed still listed as open. Those cost you the customer at the moment of decision, but they are usually the easiest to repair, because they sit in clearly structured data feeds like directories and map listings. B2B services fail differently again, most often through references attached to the wrong client or case studies that were never theirs. Work out which failure would hurt you most, and point your monitoring at that instead of watching everything equally.

Putting a number on one wrong answer

Make it concrete, and treat every figure here as a placeholder for your own. Say a mid-sized trades business is described by AI assistants as working only in its home region when it in fact delivers nationwide. Suppose 400 people a month ask an assistant about that kind of work in that area, and the wrong answer costs it 5 percent of them: 20 inquiries that never arrive. At a 10 percent close rate and an average job worth 2,500 euros, that is 5,000 euros a month it never sees and never gets to attribute to anything.

Over a year that is 60,000 euros bleeding out invisibly, because nobody connects a quiet quarter to a sentence in a chat window. Set against it: a few days of work cleaning up the sources, plus a monthly check. Two things fall out of that arithmetic. Small errors compound, because they keep answering every query, every day, until somebody fixes them. And clean data is cheap next to the revenue it protects. Substitute your own volume, close rate and deal size, and you will know in ten minutes how high on the list this belongs.

Three beliefs that waste your time

The first: if I just tell the AI the right answer once, the problem is solved. It isn't. A model doesn't learn from your chat, and correcting it in a conversation changes nothing outside that window; the underlying training data and the pages it retrieves are untouched. Correction only sticks when you change what the systems read: your own site, the directories, the official registers, Wikipedia and the databases around it, coverage in press and trade media.

The second: the louder I complain, the faster it changes. Neither a public post nor twenty reports moves this on its own. What moves it is the correct version becoming the consistent version across enough credible sources that guessing wrong stops being probable. The third belief is that this is a big-brand problem. It is usually the reverse: a small company with a thin public record gives a model more room to invent, and that is exactly where the confident nonsense comes from. Publish little about yourself and you leave the gaps for something else to fill.