What AI Gets Wrong About Your Business
The most useful thing an AI assistant will ever tell you about your business is something wrong.
Not because errors are good, but because they are actionable in a way that visibility is not. "You do not appear in this answer" leaves you guessing. "This answer says you close at 5pm and you close at 7pm" points at a specific field, in a specific source, that you can correct this week — and that is costing you calls right now.
Where the wrong information comes from
An assistant with web retrieval active is summarizing pages it just fetched. It is not inventing your hours; it is repeating something it read. So every error traces back to a source, and the job is finding which one.
The usual suspects, roughly in order of how often I find them at fault:
Your Google Business Profile. Feeds an enormous amount of what gets said about local businesses. Also the most likely to be quietly out of date — hours that were right two years ago, a service area from before you expanded, a category chosen when you started. Directories and aggregators. Yellow-page-style sites, trade directories, industry listings. Many populate themselves from other directories, so one wrong record propagates, and the original source can be a site you have never heard of. Review platforms. Usually accurate on identity, frequently stale on details. Your own website. More often than owners expect. An old phone number in a footer. A location page for a market you left. A PDF price list from 2023 still sitting at a URL nobody has linked in years but crawlers still reach. Data aggregators feeding everyone else. The wholesale layer. Fixing a downstream directory while the upstream aggregator still carries the wrong record means it comes back. Genuine confusion with a similarly named business. Common, and the most annoying to unpick.Finding out what is actually being said
Ask directly, and ask in the ways a customer would:
- "What are [business name]'s hours?"
- "What areas does [business name] serve?"
- "What services does [business name] offer?"
- "Is [business name] licensed?"
- "What is [business name]'s phone number?"
- "Tell me about [business name]"
That last point decides what you do next. If sources were cited, you have a lead. If not, the model may be answering from training data, which reflects the web as it was months ago and which you cannot edit. In that case the fix is to correct the live sources and wait; there is no faster path, and anyone offering one is describing a mechanism that does not exist.
Working backwards to the source
When an answer is wrong and sources were cited:
- Open every cited source and check whether the error is there. Usually it is, in the first or second one.
- If it is not, search the wrong value itself — the old phone number, the old address — in quotes. This finds the sites still carrying it, and it surprises people how many there are.
- Fix the authoritative source first. Your own site and your Google Business Profile, in that order.
- Then the aggregators, because they refill the directories.
- Then the directories that matter, meaning the ones that actually appeared in your citations. Ignore the rest; a service that submits to four hundred directories is selling volume, not correction.
- Record what you changed and when. You will want the date later, because propagation is slow and irregular and you will otherwise not know whether something worked.
How long correction takes
Longer than you want, and unevenly. Your own site is immediate. A Business Profile edit can be quick or can sit in review. Directories vary from days to never. Training-data-based answers change only when a model is retrained, which is not on your schedule and is not announced.
Anyone who quotes you a timeline for "fixing what AI says about you" is quoting a number they cannot know.
What you cannot correct
Be clear-eyed about the boundary:
- Opinion. A summary that says reviews are mixed, when reviews are mixed, is not an error.
- A model's parametric memory. No editing interface exists.
- Another company's content about you, beyond the ordinary routes of asking, or a legal remedy if it is defamatory.
- Whether you are mentioned at all. Correcting errors makes the record accurate. It does not compel inclusion, and nobody controls inclusion.
Keeping it from drifting again
Errors return, because the underlying sources drift.
- Re-run the questions above quarterly, same wording, same spreadsheet.
- Update every source the same day you change anything real — hours, phone, address, service area, a service you stop offering.
- Keep a written list of every place your business information lives. Most owners cannot name half of them, which is precisely why the wrong ones survive.
- Treat a holiday-hours change as a real change. Wrong holiday hours produce the most annoyed phone calls of the year.
Why this is worth an afternoon
Chasing an AI recommendation is speculative work with an uncertain payoff. Correcting a wrong phone number is not. One of these is a project; the other is a leak.
Businesses I work with are routinely surprised by what turns up — service areas they abandoned, a partner's number from a defunct arrangement, hours from before a schedule change. None of it was malicious. All of it was costing something.
Start there. It is the part of this field where the effort and the result are actually connected.
Frequently Asked Questions
Where does wrong information usually come from?
Most often a stale Google Business Profile, then directories that copy each other, then the business's own website — an old number in a footer, a page for a market it left. Data aggregators sit upstream, so fixing a directory without fixing them means it returns.
How long does a correction take to show up?
Unevenly. Your own site is immediate, a profile edit can sit in review, directories range from days to never, and answers drawn from training data change only when a model is retrained. Anyone quoting you a timeline is quoting a number they cannot know.
What cannot be corrected?
Opinion, where it is accurate. A model's internal memory, which has no editing interface. Another company's content about you, beyond ordinary routes. And whether you are mentioned at all — accuracy is not inclusion.
Do I need a service that submits to hundreds of directories?
No. Fix the sources that actually appeared in your citations. A submission service sells volume; what matters is that the records carrying you agree with each other.
Why prioritize errors over visibility?
Because the effort and the result are connected. Chasing a recommendation is speculative. A wrong phone number is a leak, and closing it is finite work with a certain payoff.