First, classify: four kinds of error, four different fixes
AI states facts from its source mix: your website, third-party platforms, Wikipedia, the press, review sites. Before doing anything, classify the error: factual errors (founding year, address, pricing), positioning errors (wrong service boundaries), misattribution (a competitor's case credited to you), or stale information (old products, old wording). The class decides the fix — factual errors mean fixing the source page, stale information means retiring the old version first, misattribution means supplying one canonical version. This page merges the response into a single four-step framework: document and classify → trace the source → fix sources and publish one canonical version → re-test to confirm.
Step 1: Document and classify
Archive the wrong answer: a screenshot, the exact prompt, the full response, the engine and the date, kept together as one record. The archive serves two purposes: it gives the re-test a like-for-like comparison, and it is the evidence you attach when filing a correction with a third-party platform or media outlet. Once archived, classify the error into one of the four kinds above and rank it by damage (see "What to fix first" below).
Step 2: Trace the source
Follow the answer's citations and the engine's knowledge sources to the polluting page — usually a third-party directory, an old press release, an encyclopedia entry, or your own outdated page feeding it wrong facts. Test the wrong claim against your site first: many errors actually originate in your own website's outdated pages, and fixing your own pages is fastest. If your site is right, work through the third-party pages in order of how much engines cite them.
For example: AI states your old office address, and tracing finds a directory profile that has not been updated for years — while your own site footer still links a PDF company profile carrying the old address. Both sources need fixing; fix one, and the next sampling round gets it wrong again.
Step 3: Fix sources, and publish one canonical version
Correct every source page: fix the third-party pages where you have edit rights (platform profiles, partner pages) one by one; for independent media, request corrections from the outlet; retire or update your own old pages the same day. At the same time, publish one authoritative fact source on your own site: key facts (services, dates, locations, credentials) stated consistently and made machine-readable (llms.txt + JSON-LD), so engines have an authoritative source to land on. For how to write an llms.txt, see What is llms.txt? Does a corporate website need one?
Asking the engine to "fix the answer" barely works — with the source unclean, the next sampling round gets it wrong again. Vendor feedback channels are an escalation route, not the main fix: submit through each vendor's official feedback channel (for example, thumbs down under a ChatGPT response with a description of the error, or the report control under a Perplexity answer) — and accept that feedback carries no promise to change any specific answer. What actually changes answers is the re-crawl after the sources are fixed.
Step 4: Re-test to confirm
Re-sample the same fixed question set that produced the error after 2–4 weeks and track the error rate separately. Fact correction follows the engines' re-crawl cycles — weeks, not days — and only sampling shows whether it is fixed. Factual accuracy is a standing metric we track for every GEO client. For the resampling method, see How do you measure a brand's visibility in AI answers?
What to fix first, by damage
Pricing and service-boundary errors (they lose deals directly) > contact and legal-entity errors (customers cannot find you) > positioning and case errors (they cost trust) > wording blemishes (schedule them). If the error lives on a platform with an edit path (encyclopedias, directories, maps), file the correction the same day; if it is your own old page, retire or update it the same day; if it lives in sourceless model memory, only sustained correct wording plus third-party corroboration overwrites it — which is why fact governance belongs before the incident, not after.
The standing work: govern facts before anything goes wrong
In our measurement of 30 global-facing websites of Chinese exporting companies (Negency Flagship Report 2026, Issue 1, sampled on a single day, 30 October 2026), 17 had no JSON-LD at all and 13 had a missing or non-unique H1 — most companies do not even have one machine-readable canonical version on their own site, so when AI gets something wrong there is nothing authoritative to fall back on. The standing work of fact governance is: keep one machine-readable fact file on the site (llms.txt + JSON-LD), keep key facts identical everywhere, and resample a fixed question set regularly to track factual accuracy. Fixing it after the incident costs far more effort than maintaining it.