Why do AI assistants get facts about your business wrong?

The instinct on seeing a wrong answer is that the system invented it. Sometimes it did. Far more often it found something — an old address on a directory, a service you dropped two years ago, hours from before you changed them, a phone number from a previous premises — and reported it accurately from a source that is out of date. The error is real and its origin is usually external and correctable.

The most common cause is inconsistency across the places that describe you. A business listed with three different phone numbers across various directories, or two spellings of its name, or an old address on one aggregator, gives a system contradictory inputs. What it produces then is a plausible reconciliation, which may be any of the versions or a blend of them, and it will state whatever it produces with the same confidence as a correct answer.

The second cause is absence, and it is the one businesses can most directly fix. If a fact about your business is not published anywhere — your service area, what you do not do, your minimum, which of two similar services you actually offer — then no source exists to draw on, and the system will either decline or generalise from businesses like yours. Generalising produces the confidently generic answer that is wrong in the particular, and the remedy is to publish the fact plainly somewhere durable.

The third is a stale primary source, which is to say your own site. Pages that were accurate three years ago, a services list that includes something discontinued, a team page with people who left, a price that has changed. Your own site is normally weighted heavily, so an outdated fact there propagates rather than being corrected by anything else. Auditing your own pages for things that are simply no longer true is unglamorous and is the highest-return correction available.

Fixing this is a sequence rather than an appeal. Correct your own site first. Then correct the major directory and platform listings, which is tedious and finite, ensuring the name, address, phone number and category match exactly across them. Then publish the facts that were missing, plainly, in text. Then re-ask the question after a period, because these systems update on their own schedule and nothing about the fix is instant.

One thing worth accepting: you will not get a correction issued. There is generally no mechanism to tell a generated answer it is wrong, and the ones that exist are limited. What you can change is the evidence available, and the evidence is the published record. That is a slower remedy than a correction request and it is more durable, because it fixes the input rather than one output.