What AI Gets Wrong About Your Business
When an AI answer gives the wrong phone number, hours or service area, start with the exact claim. This guide is a correction workflow for factual errors. For measuring how often a business appears, use the AI visibility testing worksheet.
Save the claim before changing anything
Record the answer, prompt, platform, date and source links shown. Note the correct fact and where it can be verified. Distinguish an outdated fact from a subjective opinion or an answer about a different company with the same name.
For example, “the answer lists our old phone number” is testable. “The answer did not recommend us first” is not a factual correction request.
Check whether the citation supports the claim
Open each relevant source and read the passage in context. A citation is a lead to investigate, not proof that the claim appears on that page. The source may be accurate while the answer misinterprets it, combines unrelated businesses or invents a detail.
OpenAI's reliability guidance explicitly warns that answers and references can be incorrect. Do not assume that every AI error traces to a stale web listing, or that a response without citations necessarily used no retrieval.
Use this decision table:
| What you find | Next action | Evidence of completion |
|---|---|---|
| Wrong fact on your website | Correct the page and related visible fields | Public page readback |
| Wrong fact on a profile you manage | Submit the accurate value through its editor | Accepted edit and public readback |
| Wrong fact on a third-party page | Send a focused correction request through its normal process | Publisher response or corrected page |
| Source is correct but answer is wrong | Use the AI product's feedback mechanism | Feedback receipt, if supplied |
| No supporting source identified | Record the uncertainty and check likely public records | Investigation log, not a claimed fix |
Correct the records you control
Check contact pages, footers, service pages and structured data for the same outdated value. Update visible information and matching structured data together. Do not leave an old phone number in a downloadable document that remains publicly accessible.
For a Business Profile, check the current published value as well as any pending edit. Submitting a change does not mean Google has accepted it. Keep proof of the business fact available if the platform requests verification.
Request a precise third-party correction
Identify the page, the inaccurate statement, the correct value and a public supporting source. Keep the request limited to the factual change. Do not include private customer records or account credentials.
If several directories share the same error, investigate whether they use a common data supplier. Fixing a supplier may help, but do not assume every directory automatically syncs from it. Confirm each important public record individually.
This article describes the workflow; it does not send requests or grant anyone authority to contact publishers on your behalf.
Use product feedback when the source is accurate
If the linked page is correct and the answer is not, editing that page to match the model would make the public record worse. Submit feedback with the specific inaccurate claim and supporting source where the product offers that option.
Feedback is a request for review, not a direct edit to a model. You cannot infer a successful correction from a submitted form or a more accurate answer in a single later test.
Keep three separate completion states
First, verify that the authoritative source is correct. Second, verify the important third-party records that you can inspect. Third, retest the original question and record what the AI product says at that time.
These states may complete at different times. Source publication, crawling, retrieval and answer generation are separate processes. There is no reliable universal deadline for a correction to appear in every AI response.
Maintain a correction register
Record the wrong value, correct value, source URL, owner, action date and public verification date. Add a later observation column for the AI response. Mark unresolved items explicitly.
Prioritize errors that could misdirect a customer, such as contact details or service availability. Accurate records are the immediate deliverable. A ranking improvement, recommendation or new customer requires separate evidence.