- Treat it like any data quality problem: record it, trace it, fix the source, retest.
- Most errors come from old pages, inconsistent listings or thin information, not from malice.
- You cannot edit a model directly. You can change what it reads.
It is an uncomfortable moment. You ask an AI engine about your company and it gives the wrong founding year, an outdated service, a price you never charged or a competitor’s address. The instinct is to complain. A better instinct is to treat it as a debugging task.
Step 1: record exactly what it said
Capture the prompt, the engine, the date, the location and the full answer. Run it three times. If the error appears in all three, it is stable and worth fixing. If it appears once, note it and move on.
Step 2: find where it might have come from
- Your own old pages, such as a service you dropped or a price page from last year.
- Business directories with stale details.
- A news article or press release that has since changed.
- A competitor or reseller page that mentions you incorrectly.
- Your own site contradicting itself between pages.
If the engine shows sources, start there. If it does not, search the wrong claim in quotes in Google. You will often find the exact page.
Prompt, engine, date, answer.
Find the page that says it.
Fix or request a fix at source.
Rerun in a few weeks.
Step 3: correct what you control
Update or remove the wrong page. Do not leave an old price page live for the sake of tradition. If you must keep it, mark it clearly as archived and point to the current one. Make sure the facts are stated in plain text, not only inside images or PDFs. Then align your structured data with the visible page.
Step 4: correct what you do not control
- Ask directory owners to update the listing, or use the claim and edit tools they provide.
- Contact the publisher of an outdated article and ask for a correction, with the evidence.
- Publish one clear page that states the correct fact, such as a company facts page, so there is a reliable source to cite.
Step 5: tell the engine, where there is a route
Most engines offer a feedback control on answers. Use it, with a short factual note. Do not expect an instant change, and do not rely on it alone.
Step 6: retest, and be patient
Rerun the same prompts after a few weeks. Changes to search based answers can appear as pages are recrawled, which Google says can take days to months. Changes to what a model learned in training depend on the next update and are outside your control.
Prevent the next one
Keep a one page facts sheet with your legal name, services, locations, founding year and pricing approach, and keep it current. Every directory, profile and press kit should copy from it. Consistency is the cheapest defence. The same idea runs through our GEO work, and the checks are described in our audit guide.