Why AI sometimes gets your business information wrong
When an AI assistant gets your business wrong, the error almost always lands in one of three buckets: stale (true once, not anymore), conflicting (your own sources disagree), or invented (the model filled a gap with something plausible). The buckets matter because each one has a different cause and a different fix. Diagnose first, then act.
The three failure modes
Stale means the AI is repeating information that used to be true: your old address, your old hours, your old prices. Conflicting means your public sources disagree and the AI picked the wrong one, or blended them. Invented means the model generated a plausible detail with no source at all. Nearly every wrong answer about a local business is one of these three.
Stale: the answer that’s two years old
Models carry a memory with a documented end date. AI companies publish these dates: Anthropic’s model documentation, for instance, lists the exact training cutoff for each of its models1. Anything you changed after that date doesn’t exist in the model’s memory, so an answer drawn from memory can recite your 2024 self with full confidence. The training pile updates on the vendor’s schedule, not yours, which is why staleness is the one failure mode you partly wait out. Partly: live lookups read your current sources, so keeping those current is how you stop staleness from spreading into fresh answers.
Conflicting: when your listings disagree
The dead phone number on a directory you forgot. The old suite number on your profile. Hours that differ between your site and your listings because one got updated after the holidays and the other didn’t. A machine reading five sources that disagree has to pick, and it doesn’t know which one you meant. This failure mode is entirely yours to prevent, it’s the cheapest one to fix, and consistency is a signal in its own right, not just error prevention. It’s also the mode that most often turns into skipping you entirely: conflicting sources make a business risky to recommend, and a machine with other candidates just moves on.
Invented: hallucination, without the hype
Models do sometimes generate details no source contains: a service you don’t offer, a founding year, a warranty policy. Honest sizing, without numbers we can’t support: invention shows up most where your public information is thinnest, because the model is completing a pattern with nothing to check against. A business with rich, consistent public information gives the model less blank space to fill. That’s the practical takeaway hiding under the hallucination headlines: unreliability about your business is partly a signal problem, and the signal is yours. What invention is not: fixable by complaint alone. There’s no customer service line inside a model.
What you can fix and what you can only monitor
Fixable: everything the assistant reads live. Update the profile fields AI actually pulls from, correct the listings, make your site say the current truth plainly, and the lookup pipeline starts serving the corrected version in days to weeks. The per-platform fix paths are their own walkthrough, including the feedback mechanisms and what to realistically expect from them.
Monitorable only: the memory side. You can’t edit a trained model, and our opinion is that you shouldn’t spend a dollar trying; anyone selling “hallucination removal” is selling a service they can’t perform. Watch what the answers say about you on a rhythm, fix the live sources when something’s wrong, and let the model versions roll forward on their own clock.