Absence is the problem owners expect. Misinformation is the one that costs more, and it is quieter, because nobody tells you it is happening.
A prospective customer asks an assistant about you — by then they know your name, which means they are close to buying — and reads that you closed during the pandemic, or that your prices start at a figure you abandoned three years ago, or that you do not offer the service that is now most of your revenue. They do not call to check. They cross you off and ask the assistant what else is available.
Where wrong facts come from
Assistants do not invent business facts from nothing very often. Usually they are repeating something, and the something is findable.
Stale listings. A directory profile created in 2019 and never revisited. An old address, a closed branch, a discontinued service, a phone number from two premises ago. These pages are structured and machine-readable, which is exactly why they carry weight — and exactly why the error is stated with such confidence.
Superseded pages on your own site. A pricing page still live at an old URL, a services page you replaced but never removed, a landing page from a campaign you stopped running. If it is reachable, it is quotable.
Genuine confusion between businesses. A similar name in a nearby city, a former partner who kept a similar trading name, a franchise that shares your brand but not your ownership. Models blend entities that look alike in text, and the blend can be quite specific: your address with their opening hours.
Old news, permanently fresh. A closure, a change of ownership, a now-resolved dispute. A page from 2021 that was accurate in 2021 is not marked as historical anywhere a model can see, so it goes on being repeated as though it were current.
Aged inference. Occasionally a model states something plausible and unsourced — a certification, a size, a specialism — assembled from patterns rather than from any page. These are the hardest to trace and the ones most worth catching, because there is nothing to correct; there is only material to outweigh.
Why arguing with the assistant does not work
The instinct is to tell it. You correct it in the chat, it apologises gracefully, restates the fact correctly, and you feel the matter is settled.
Nothing was settled. That correction lived in that conversation. The next customer opens a fresh chat with no memory of it, asks the same question, and gets the same wrong answer. Assistants do not learn from being contradicted by a stranger; if they did, the failure mode would be far worse than a wrong price.
The repair has to happen where the claim lives. That is generally a page, and pages can be edited, replaced, or outnumbered.
The order of work
- Establish exactly what is being said, and by which assistants. Not approximately. The specific wording, from a fresh chat, dated, recorded. You will need it to know whether anything you do afterwards worked.
- Find the source. Where citations are shown, read them. Where they are not, search for the wrong fact itself — the old price, the closed address — and see which page still carries it. It is usually the first result.
- Correct at the source. Update the listing. Remove or redirect the stale page on your own site. Contact the author of the article. Unglamorous, and generally quick once you know which page to touch.
- Make the correct version easy to find and hard to misread. Stated plainly, in one place, in language that can be lifted: what you do, where, for whom, at what price, with what qualifications. Ambiguity is what let the wrong version win in the first place.
- Re-check on a schedule. Live-search assistants pick up corrections within days to weeks of a recrawl. Facts absorbed into training data may not shift until that model is retrained — which is why step 4 matters. You are not only deleting the old claim; you are ensuring that when the model next learns about you, the consistent version is the one available.
The part that is genuinely uncomfortable
You cannot fix what you have not read, and nobody is going to forward you a screenshot of an assistant getting you wrong. The customer who saw it is gone and has no reason to tell you. The error can run for a year, across four assistants and every phrasing of every question, and produce no signal at all except an inexplicably lower conversion rate on people who seemed keen.
That asymmetry is what makes this worth checking deliberately rather than waiting to notice. Absence at least matches the intuition that you need more marketing. A confident, wrong answer looks exactly like nothing happening.