# How to Track Wrong Business Facts in AI Answers By AI Visibility editorial team AI-assisted article. An AI answer gives customers the wrong pickup time for a Dubai branch. The branch manager confirms the time, the web team updates an old location page, and a fresh answer gives the right information. Three weeks later, someone gets the original wrong time again. If the first ticket simply says “fixed,” the team has to reconstruct what it saw, which page changed and what the follow-up answer actually said. Keep a factual error log that separates three things. The observation records the answer customers could see. The issue records the business fact that needs checking. The correction record shows which source changed and what subsequent observations found. That separation lets you reopen a recurring error without losing the history and lets your monthly report distinguish source maintenance from changes in AI answers. Capture the answer before opening a correction ticket Save the exact passage, including the branch, date or qualification surrounding the statement. Record the question, answer surface, time, language, region and whether the answer followed earlier messages. If a source panel was available, preserve its link list and any evidence that a particular link was attached to the statement. A citation listed somewhere in the answer is weaker evidence of a specific claim's origin than a link attached to that claim. Use a screenshot or answer export to preserve context where the capture process allows it, and keep the business-facing summary in a separate field. Split statements into facts you can verify independently. “The Marina branch takes walk-ins until 9 p.m.” can contain distinct assertions about the location, walk-in policy and hours. Check which part the answer attaches the time to. If 9 p.m. is the store closing time and the answer applies it to appointments, the issue is a mismatch between service hours and store hours. A single vague “wrong opening hours” ticket can send the owner to edit the correct field. Give each observation an ID. Each issue needs its own ID, even when both initially live in one spreadsheet. An answer can contain several issue observations, and several answers can repeat one issue. Link the records and keep the original answer unchanged. If someone later realizes that a screenshot belongs to a different branch or date, add the review decision and its evidence. This preserves the reason the issue was first raised and the reason its classification changed. Establish what was true for that branch and date For every fact, identify the entity, location or service, field, value, and dates when the value applied. The business's current website may be an appropriate public reference, but use the correct page and obtain confirmation from the team responsible for the fact if sources disagree. An operations lead can confirm the actual opening schedule. The booking team may be responsible for service availability. Record who confirmed the fact, what they confirmed and when the team checked it. Keep the effective date separate from the verification date. A post dated Thursday can announce a schedule that starts on Saturday. A later change to the regular schedule does not make an older answer incorrect if the older value applied when the answer was captured. On the other hand, a question about a future booking date should be judged against the policy in effect for that booking date if reliable evidence establishes it. Record which date the customer asked about. Assign a verdict to the specific claim. A confirmed error has evidence that contradicts the claim within scope. A correct claim has supporting evidence within scope. An unresolved claim has insufficient or conflicting evidence. A claim outside the audit scope can be retained separately, as can a question that received no relevant claim. For example, an answer that gives no pickup time is a missing value in this audit. An answer that says there is no pickup service makes an assertion that can be verified. Google Business Profile documentation separates regular hours from special hours and hours for particular services. That distinction is useful when you construct the fact register. It lets a reviewer check whether the business's ordinary schedule, a holiday exception or a service-specific schedule applies. The Google field definitions describe what you can record on the profile. A comparison with an answer still needs the capture and date-specific evidence. Google Business Profile hours documentation (https://support.google.com/business/answer/15300403) Keep publisher evidence within its scope A public source can establish that a business made a statement without settling whether the statement applies to the audited location or service. Save the exact URL and the date reviewed. For a social post, check who published it, read the offer or service description and retain the dates and regional limitations stated there. For a branch page, identify the address and any service-specific conditions. When relying on a business's internal confirmation, separate the private evidence from the public link that a reader can inspect. The official Urban Company UAE Instagram example describes sofa cleaning. It can document the scope of that promotion and its publisher. If an audit concerns a different service or a specific appointment time, the team needs evidence for that service or time. Keep marketing material in the record only for the assertions it actually supports. Record evidence independently of any theory about how the wrong answer arose. A stale third-party listing with the same incorrect value is a plausible correction target. A matching value alone does not prove that the listing supplied the answer. You can record “matching stale listing found” and separately record “source association unknown.” In practical terms, the web team can fix the listing without having to solve how the answer was generated. W3C's provenance work describes how records can connect the things produced, the activities that produced them and the people or organizations involved. A business spreadsheet can use that distinction with separate fields for an answer capture, a verification event and a responsible reviewer. An evidence trail remains usable when a correction is later challenged or reassigned. W3C PROV overview (https://www.w3.org/TR/prov-overview/) View the selected official Instagram post (https://www.instagram.com/urbancompany_uae/reel/DfCnVfrJsBc/) Official Urban Company UAE sofa-cleaning promotion, previously verified. Its caption illustrates scope of publisher evidence. Give each issue an owner and a next check A useful issue key combines the location or entity, the fact field, and its scope. Include the dates when the scope is temporary. Multiple answers saying a branch offers the same unavailable service can link to one issue even if they phrase it differently. A subsequent wrong claim about a different branch belongs on its own issue. A new seasonal schedule can also call for a separate issue because the dates change the evidence and closure decision. Assign a responsible person and a due date after triage. Prioritize what could make a customer take the wrong action. A false “walk-ins accepted” claim can send someone to a location that cannot help them. A wrong booking link may cause a customer to contact an unrelated provider. A small spelling difference may have little practical effect if the answer points to the right place. Record the impact judgment and its basis separately from the observed facts. If there are actual customer complaints, the record should show where they came from and avoid exposing personal details. Track source actions separately from issue status. A source action can be requested, published or independently verified. For example, “profile update submitted” is evidence that someone filed the change. “Public listing shows new hours” is evidence that the profile currently displays the corrected value. Each action should carry a link or other reviewable record, effective date if relevant, and an owner. That detail helps when one website is corrected while a directory still shows the old value. For the answer itself, use clear status transitions. A new issue can move from unverified to confirmed open, then to monitoring after a documented correction or investigation. Record retest outcomes as events. A closure rule can require enough relevant answers to check the original assertion again, and the team can reopen an issue if the same wrong assertion returns. When the period ends or the service is retired, resolve the monitoring task with an explicit reason and keep the historical error count intact. Read the correction timeline one event at a time The following example is constructed. The fictitious Marina branch offers pickup until 6 p.m. on ordinary Thursdays. The storefront stays open until 8 p.m. The audit question asks when customers can collect an order from that branch on a normal Thursday. The web team maintains a pickup page and the operations team owns the hours. Event What the reviewer records What the issue record says First answer says pickup ends at 8 p.m. The passage and dated pickup schedule contradict each other Confirmed open Pickup page is revised and checked The public page now distinguishes pickup and storefront hours Monitoring First follow-up gives a correct 6 p.m. cutoff Correct relevant assertion under the comparable question Monitoring Second follow-up gives no pickup hours No relevant assertion available to check Monitoring Third follow-up again says 8 p.m. Another confirmed observation linked to the same issue Confirmed open The last event adds a repeat occurrence. It does not create a second underlying issue. The public page revision remains verified, even though the answer repeated the old value. Likewise, the missing hours in the second follow-up cannot count toward a closure requirement for correct relevant assertions. Someone reading the log can see that the team made a documented change, observed one correct result and later saw the error recur. If you need a figure for the follow-up, there were three completed answer captures, two with an adjudicated pickup-time assertion, and one confirmed wrong assertion among those two. The third capture provides evidence of omission. Describe that result as a small worked tally. A different question set, surface or period might produce different results. Save the original question so reviewers can compare like with like when the issue returns. Put the issue inventory beside the error rate A monthly report needs an inventory and a rate with a named unit. Show distinct confirmed issues opened, issues still open at the cutoff, and issues reopened after monitoring or closure. Explain whether new issues mean first seen that month or first confirmed that month. If the verification team is behind, those dates can fall in different periods. Use an explicit month-end cutoff for the inventory and retain dated status changes so earlier reports can be reproduced. For the rate, a straightforward definition is confirmed wrong assertions divided by the count of adjudicated assertions within scope. State how many assertions were unresolved and how many completed captures gave no relevant assertion. If you want to know how often customers saw any wrong fact, calculate a separate answer-level rate with its own denominator. A single answer can contain several incorrect facts and several correct ones. Counting that answer once addresses a different question from counting each assertion. For correction effectiveness, report the retest population and show how much it covers. You can say “four issues had relevant follow-up observations” and separately list the outcomes. A publication count belongs beside these figures, since it measures work completed on sources. A month with many corrections published and few comparable retests has evidence of maintenance progress. It has limited evidence about the current answer experience. The monthly report can show both without combining them into one score. AI Visibility monthly reporting guide (/article/what-should-a-monthly-ai-visibility-report-contain/) Factual error log record Capture ID, timestamp and conditions, exact assertion, entity and branch, fact field, relevant date, verdict and evidence, issue ID, assigned owner, next check, source action and verification status, and every follow-up observation. Keep the closure criterion next to the issue and record the reason for any exception. That makes it possible for a second reviewer to explain why a ticket closed, why it reopened and which facts a customer could actually see. AI Visibility is published by Lunasol, which offers GEO and AEO services. The pickup example is constructed to illustrate the method. About AI Visibility (https://aivisibility.ae/about/)