AI answer engines state facts with confidence, and sometimes those facts are wrong: an outdated fee, a feature you do not offer, a guarantee you never made. A confident wrong answer is worse than silence, because a buyer acts on it and, in a regulated category, a regulator may read it too. Accuracy is a measurable dimension of AI visibility, and misrepresentation is a distinct state you can find and fix, not an unavoidable cost of being discovered.
Why engines get brands wrong
Models blend what they learned in training with what they retrieve live, and both can be stale or thin. If the strongest source about a detail is an old page, a competitor's comparison, or a forum thread with a mistake, the engine can repeat it with total confidence. It is not lying; it is faithfully summarising a flawed source.
The thinner your own authoritative footprint on a fact, the more the engine leans on whatever else it can find. Gaps get filled by the loudest available source, which is often not you.
Misrepresentation is its own visibility state
It is tempting to file 'wrong' under 'mentioned', but they are different. Absent is invisible. Linked is present and correct. Misrepresented is present and wrong, and it is the most urgent state because the answer is actively working against you while looking authoritative.
Separating misrepresentation from ordinary presence changes the priority order. A misstated guarantee on a high-intent question outranks a missing mention on a low-intent one, because the first is doing damage and the second is only a lost opportunity.
How to measure accuracy
Accuracy is not a vibe; it is a comparison. Keep a set of approved, source-backed facts about your brand: your real fees, terms, coverage, and claims. Sample what engines say about you across your tracked questions and check each statement against that set. What does not match is a flagged inaccuracy with a receipt.
This only works if your facts are governed. Approved facts with sources are the reference the check runs against, and they double as the grounded material your own content and any AI-answer content should be built from. Without that reference, accuracy is one person's opinion against the model's.
Fix it at the source, then re-measure
You cannot edit the model, so you correct what it reads. Publish or update your own authoritative page on the fact, get the third-party sources the engine cites to carry the right version, and make the correct statement easy to extract and cite. The answer follows the sources over time.
Then confirm it. Re-measure the affected questions after the sources change and watch the misrepresentation flags clear. Fixing accuracy without re-measuring is hoping; fixing it and watching the flags drop is proof, which matters most when compliance asks whether the wrong claim is gone.