Why ChatGPT Has the Wrong Phone or City

Assistants assemble a name from one set of pages and a business card from licensed feeds. When those sources disagree, the model invents a bridge. A dated public file gives both jobs the same facts: legal name, website the owner controls, phone, and city or service area.

Published 11 September 2026

Wrong AI information is usually not malice. It is two trucks arriving with different cargo. Editorial pages, listicles, and forums supply the sentence that names a firm. Map listings and licensed local feeds supply the card beside the sentence: address, hours, phone. If the website says one trading name, the map listing says another, and last year’s directory still has the old number, the model has to guess.

Wikidata deletion makes the guess worse. The knowledge-base item that could have anchored the name is gone. The leftover pages still disagree.

Fix the sources, then publish a file

Fix the four facts on every surface you control. Then publish a third-party file that repeats them and dates them. Put a seal on the site that points at that file so a crawler does not have to finish the theme to reach the same object.

Gold is $297 a year after website proof. It will not force every answer product to recite the file. It gives the products one object they can fetch when they look.

Why AI gets a business wrong · How to correct wrong AI information · NAP consistency · Wikidata deleted my business

Can I submit a correction to ChatGPT?
You can mark an answer as wrong. That is weak. Fix the sources the model can fetch.

Is schema enough?
Schema on your site is first-party. Use it. Add a third-party public identity file.

Does llms.txt fix a wrong phone?
Only if the phone on the pages it points to is already right. llms.txt is a map, not a second source.