How to fix a wrong AI answer about your company

A buyer asks an assistant what your product costs. It answers confidently, cites your own pricing page, and quotes a number you retired eighteen months ago. Nobody clicks through. Nobody emails to check. You lose the deal without ever learning it existed.

This is the ordinary failure of AI answers, and it is not the one most tools look for. They ask whether you were mentioned and whether the tone was positive. A confident, well-formatted, correctly cited, wrong answer passes both tests.

Why assistants get your company wrong

There are four distinct causes, and they have four different fixes. Treating them as one problem is why most correction efforts stall.

That last row is the one teams skip. If an answer was produced without grounding, fixing your robots rules for a retrieval crawler will not move it, because no retrieval happened. Knowing which mode produced the answer is a prerequisite for choosing the fix, not a detail.

How to find all of them, not only the one you spotted

The screenshot someone forwards you is a sample of one. Here is the method that produces something you can act on.

1. Write down what is true, with dates

A claim is only wrong relative to something. Before measuring anything, record your facts with an effective date and an expiry: price, fees, limits, integrations, availability, certifications, leadership. The dates matter more than they look. Most wrong answers are not fabrications, they are facts that expired, and you cannot detect an expired fact without knowing when it stopped being true.

2. Ask the questions your buyers actually ask

Not invented prompts. Pull them from search console, site search, support chat, sales calls and loss reasons. Keep them in intent families and never average across families: being named when someone asks about you by name is a different result from being recommended when someone asks for a vendor, and blending the two produces a flattering number that means nothing.

3. Ask repeatedly, and record the exact setup

"ChatGPT" is not a measurement surface. Provider, model, model version, access mode, grounding mode, search mode, country, language and personalization state all change the answer. Store all of them with every run, or you will not be able to reproduce a finding or tell which surface a fix affected.

4. Check the claim and the citation separately

Two different checks. First, does the extracted claim contradict a fact in your registry, given the dates that fact was in force. Second, does the page the answer cited actually contain the claim attached to it. A sourced-but-unsupported claim is a distinct defect with a distinct fix, and it is the one that most damages trust when a buyer follows the link.

5. Fix the record, then prove the answer moved

Correct the page, wait for the retrieval crawler, then re-ask the same questions on the same surfaces and compare against a set of questions you deliberately left alone. Without that untouched control you cannot separate your fix from a model update that happened the same week. This is the step almost everyone skips, and it is the only one that turns a correction into evidence.

What "fixed" has to mean

A number going up is not evidence. Models update, retrieval indexes shift, and a competitor publishing something unrelated can move your rate. The claim "our fix worked" requires a before window, an after window, matched controls that were not touched, and a stated effect size with its uncertainty.

It also requires enough samples to see the change at all. Detecting a ten-point move at a 40% base rate needs roughly 388 runs per side at 80% power. If you sampled thirty prompts, a ten-point improvement is invisible to you and so is a ten-point regression. We worked that arithmetic through in how many prompts an AI visibility number actually needs.

The order that matters

1. Fix the pages you own first. Highest success rate, lowest cost, fastest feedback.
2. Reconcile pages that contradict each other before writing anything new.
3. Repair retrieval access only for the crawler class that actually produced the answers you care about.
4. Send correction packets to third parties last. They are the least likely to land and the slowest to verify.

Nobody outside a frontier lab decides what a model says. What you can do is make the record it reads correct, and then measure honestly whether the answers changed. That is the whole job.