The difference between a wrong link and a wrong sentence
A buyer who finds an outdated page in a list of search results still sees the page, the date and the other nine results. They can judge. A buyer who asks an assistant and is told "that firm mostly handles residential work and covers the inner suburbs" gets a single, fluent, apparently authoritative sentence with no visible seams.
That is the whole problem. The confident register of a generated answer strips away the signals people normally use to discount information. Nothing in the phrasing tells the reader that the claim came from a three-year-old directory listing, or from a page you deleted, or from a different business with a similar name.
Assistants do sometimes hedge, and some cite their sources. But hedging is applied unevenly, citations are not always offered, and a buyer in a hurry reads the summary rather than the footnotes. In practice you should assume the description is received as fact.
The errors that actually occur
Inaccuracy in this context is rarely dramatic invention. It is mundane and specific, which is precisely why it passes unchallenged.
- Services you no longer offer, or never offered — often drawn from an old page or an inherited listing.
- A service area that is too narrow, too broad or simply wrong, usually because your own pages and third-party listings disagree.
- Confusion with a similarly named business, sometimes in another industry or another city.
- Outdated pricing or packaging, stated as current.
- Stale status: described as closed, merged or dormant when you are trading normally.
- Missing credentials — licences, certifications or specialisms omitted because they are not stated anywhere a system can read.
- Wrong size or positioning: a specialist described as a generalist, or a commercial firm framed as a domestic one.
What the error costs you
The cost is an enquiry you never hear about. A buyer told you do not cover their area does not call to check; they call whoever was named as covering it. There is no bounce to see in your analytics, no form abandoned, no record of the conversation. Accuracy failures are invisible by construction, which is why they tend to persist for years.
There is a second cost once an enquiry does arrive. If a prospect brings expectations set by an inaccurate description — a price that has moved, a service you discontinued — the conversation opens with a correction. That is a poor start, and it falls to the person least equipped to explain where the claim came from.
And there is a compounding effect. If an inaccurate claim is also the most quotable statement about you available, it gets repeated and eventually corroborated by other low-quality sources. Correcting it later means correcting the copies as well.
Omission deserves its own mention, because it is the error people notice least. A description that is technically correct but leaves out the licence, the accreditation or the specialism you compete on is not neutral — it is a weaker case for choosing you than the one you would make yourself.
Why the errors happen
Almost always for one of four reasons, and three of them are within your control.
- Stale sources. Old pages, archived content, abandoned profiles and directory entries nobody has updated remain readable long after they stopped being true.
- Self-contradiction. Your own pages, listings and profiles describing your services, coverage or trading name differently gives a system a choice, and it may not choose the current version.
- Thin information. Where specifics are absent, a system fills the gap with what is typical for your category — and your distinguishing features are exactly what gets lost.
- Name collision. A shared or similar name merges two businesses into one description, which no amount of your own content will resolve unless you state distinguishing detail explicitly.
How to correct it, realistically
You cannot edit a model, file a correction request, or have a statement retracted. What you can do is change the evidence, which works but is slower than people expect.
Start by finding the error precisely: ask the assistant to describe your business, several times, on each platform that matters, and log each incorrect claim as its own item. Then trace where that claim plausibly comes from — an old page, a third-party listing, a profile you forgot — and correct or remove it at the source, rather than only adding a contradiction elsewhere.
Next, make the correct version unmissable: state it plainly on your own site, including the negative form where it helps ("we do not undertake residential work"), and align every third-party listing to the same wording. Then wait, and re-check on a schedule. Retrieval-based answers often update within weeks; anything sitting in a model’s internal knowledge can take considerably longer, and no platform publishes a timeline.
Treat accuracy as a monitored metric
Because errors are silent, the only reliable defence is periodic checking. Add a small set of brand-description prompts to whatever measurement routine you run, log every factual claim, and track the share of answers containing no error. That is an accuracy rate, and it is the one AI metric that is fully about facts you already know.
The method sits alongside the broader approach in our article on measuring AI visibility. If you are still establishing whether you appear at all, start with what AI visibility is, because presence and accuracy are genuinely separate problems with separate fixes.
Give the check an owner and a date in the calendar. Accuracy drifts every time a service changes, an office moves or a qualification lapses, and the business rarely notices because the drift happens in sources nobody on the team has any reason to look at.
Last reviewed 7 August 2026. AI Says is not affiliated with or endorsed by OpenAI, Google, Anthropic, Perplexity or other AI and search providers; platform names are used descriptively.