AI Hallucinations About Brands: How to Find and Correct Wrong Facts in ChatGPT Answers
This guide was produced by AI, TELL ME!, a full-cycle AEO & GEO agency with our own AI search monitoring platform.
Summary
A ChatGPT answer can name a brand, cite it and leave the sentence about it unchecked by anyone, and an August 2026 audit by Empirank put numbers on that gap: of 1,257 testable claims about local businesses, 24 were contradicted by evidence and 479 could not be checked either way. AI hallucinations about brands therefore split into errors that need a correction and gaps that need public evidence, and the second group is about twenty times larger. Language models also struggle with less popular facts, which leaves smaller brands with the most exposure. The sections below sort claims into four states, show which facts break first and set out how to audit an answer claim by claim and fix the sources behind it.

Visibility reports count how often a brand appears in answers from LLMs (large language models) and say little about what those answers state. A ChatGPT response can name a company first, link to its site and attach a credential that no public page documents. AI hallucinations about brands matter because a buyer reads the sentence around the mention.
Sorting comes before fixing, since a claim can be wrong, unverifiable or fine and each state has a different remedy. The sections below use an August 2026 audit that keeps those states apart, two academic studies on why models err and a six-step check that ends in source corrections.
How common are AI hallucinations about brands?
In an August 2026 audit, 1.9% of testable claims about recommended local businesses were incorrect, and 11.0% of the businesses carried at least one incorrect claim. Empirank, a research firm working on answer engine optimisation (AEO), published the audit on 18 August 2026. Its sample started with 3,000 local businesses and narrowed to the 182 that a search-enabled OpenAI configuration recommended and whose identity could be resolved safely. Those 182 businesses produced 1,257 factual claims that evidence could test, and PPC Land's report on the audit sets out the full breakdown.
At claim level, 24 of 1,257 claims were contradicted by the evidence, which looks small. At business level, 20 of 182 companies carried at least one such claim, which is one in nine, and five of them carried an error the audit rated high severity. A brand lives with the business-level number, since a buyer meets one answer about one company.
The audit has limits: the responses came from one OpenAI configuration in August 2026, the businesses were local, and independent AI agents ran the quality checks with no human reviewers. Empirank notes that results may shift across models, interfaces and time, and none of the figures describe other models or other months.
Four labels for a claim, and what each one needs
Every testable claim in an AI answer about a brand lands in one of four states, and each state calls for a different fix. Empirank labelled each claim as confirmed correct, unsupported, incorrect or ambiguous, and the table below shows how the 1,257 claims split.
State | Meaning | Share in the audit | What the brand needs |
Confirmed correct | Public evidence directly supports the claim | 59.6% (749 claims) | Current records that stay consistent |
Unsupported | Evidence neither confirms nor contradicts the claim | 38.1% (479 claims) | Public evidence about the brand itself |
Incorrect | Evidence contradicts the claim | 1.9% (24 claims) | A correction at the source |
Ambiguous | Wording or evidence does not allow a decision | 0.4% (5 claims) | A clearer statement of the fact |
Unsupported claims outnumbered incorrect ones by about 20 to 1, with 479 against 24. Empirank keeps the two groups apart on purpose: adding them together gives an error rate above 40%, which would overstate model error and hide the difference between a brand that needs a correction and a brand that needs better public evidence about itself. Among the 773 claims the evidence settled either way, 96.9% were correct.
Accuracy needs a check of its own next to the five AI visibility metrics. Sentiment scores tone, and a sentence can read as neutral in tone and still be false.
Why smaller brands carry more exposure to wrong claims
LLMs handle popular entities better than obscure ones, and a brand with thin coverage sits on the obscure side. A 2022 arXiv study by Alex Mallen and colleagues tested 10 language models on PopQA, a set of 14,000 questions, and found that models struggle with less popular factual knowledge and that scaling up did little for long-tail facts. Models that retrieved documents beat models orders of magnitude larger that did not. The study dates from 2022 and was revised in 2023, so it explains a mechanism and gives no current error rate.
In three GEO studies by AI, TELL ME!, brand-owned websites supplied between 5.6% and 32.0% of the citations or citation weight behind AI answers, and other parties supplied the rest, which puts most evidence about a brand on pages the brand does not control.
The article on answer engine optimisation for education providers describes one form of the gap: a cohort-based programme described as self-paced because its pages never state the schedule, so the model fills the space from competitor pages. Fewer pages stating a fact means fewer places for a model to check it.
Which facts hold up and which break first
In the Empirank audit, identity claims held up best, with 99.5% confirmed correct, and accreditation claims held up worst, with 19.2% confirmed. The table sets the groups side by side.
Claim group | Confirmed correct | Claims tested | Reading |
Identity claims | 99.5% | 182 | Most reliable group in the audit |
Accreditation claims | 19.2% | 26 | Low mostly because no accessible page documents the credential |
Claims backed by first-party pages only | 75.2% | 105 | Descriptive comparison, groups differ in size and content |
Claims backed by third-party pages only | 57.5% | 1,064 | Descriptive comparison, groups differ in size and content |
Accreditation shows the pattern most sharply: a credential that no accessible page documents cannot be confirmed, so the audit could not call it true or false. PPC Land's report observes that trade licences, professional memberships and industry awards tend to live on registries, association directories or nowhere public.
Empirank's order of work puts location, availability, credentials, service scope and branches first, because errors there change what a customer does.
A citation does not guarantee accuracy: recommendations with an attributable citation had 59.9% of their claims confirmed (1,199 claims across 170 businesses), against 53.4% for recommendations without one (58 claims across 12 businesses). Empirank treats the 6.5-point gap as descriptive and makes no causal claim, and the uncited group was too small for a firm comparison. In both groups, 40% or more of claims stayed unconfirmed.
How to check what ChatGPT says about a brand, claim by claim
A claim-by-claim check splits each answer into single facts and labels every fact against public evidence. The FActScore paper by Sewon Min and colleagues set out the method in 2023: break a generated text into atomic facts and compute the share that a reliable source supports. ChatGPT scored 58% on generated biographies in its human evaluation. That figure belongs to 2023 models and says nothing about current ones, and the method carries over to answers about brands without change.
A brand runs the check in six steps.
Write a fact sheet of the claims a buyer would act on, such as where the brand operates, what it offers and which credentials it holds, each paired with the public page that proves it.
Run category prompts and branded prompts at least three times each, using the prompt types from the guide to checking whether ChatGPT recommends a brand, since answers vary between runs.
Split every answer into single claims and label each one confirmed, unsupported, incorrect or ambiguous.
Rank the incorrect claims by what they change for a buyer, starting with location, availability, credentials, service scope and branches.
Trace each incorrect or unsupported claim to the pages a model most likely read, starting with the sources the answer itself cites, then directories, review platforms and older press coverage.
Repeat the same prompts after each source fix and relabel the claims.
Step 5 makes the rest actionable, because a wrong claim with no traced source has nowhere to be corrected.
How do you fix wrong information about a brand in ChatGPT?
A wrong claim gets fixed at the sources a model reads, since a brand cannot edit an AI answer directly. The work runs through three kinds of source and ends with a re-check.
On the brand's own pages
Each commercial fact belongs in a plain sentence on the page that owns it, with the same wording and numbers wherever it repeats. That covers location, service scope, availability and credentials, with each credential naming and linking its issuing body. In the Empirank audit, 75.2% of claims backed by first-party pages only were confirmed, a comparison Empirank flags as descriptive and no proof that owned pages produce accurate answers. The guide to getting a website cited by ChatGPT covers page structure.
In registries, directories and review platforms
Credentials and licences tend to live on registries and association directories, and the 19.2% confirmation rate for accreditation claims points there. Each listing should repeat the same name, location, service scope and credentials as the brand's own pages, and a credential with no public listing needs one before a model can confirm it.
In earned coverage
Trade press that states the correct fact gives a model a second source to check the brand's own pages against. The article on how earned media and citations influence AI recommendations covers that route in detail.
Re-running the same prompts
The Empirank audit does not measure how quickly a corrected source changes an AI answer. Its suggested retest after source corrections is a proposal with no reported result behind it, and each model refreshes on its own schedule. A fixed re-run cycle on an unchanged prompt list is the way to see whether a fix landed.
Where an agency fits in a correction programme
An agency fits the stages that repeat in a correction programme: monitoring the answers, diagnosing the sources behind wrong and unsupported claims, setting the order of fixes, carrying them out and measuring the result. AI, TELL ME! is a full-cycle GEO (generative engine optimisation) and AEO agency with 17+ years in marketing, SEO and PR, and works through those stages with its own monitoring platform as the instrument for diagnosis and measurement.
Delivery runs beside an in-house team through Tailored GEO Consulting and GEO Workshops or as a fully outsourced function. PR is part of the offer as one extra: Third Party GEO works on editorial coverage, expert commentary, directories, rankings and reviews, which are the pages that give an unsupported claim its evidence. A team with analyst capacity can run the six steps above itself and bring in the agency for the fixes they produce.
FAQ
What are AI hallucinations about brands?
AI hallucinations about brands are statements in an AI answer that give false or invented facts about a company, such as a wrong location or a credential it does not hold. A related group, unsupported claims, cannot be checked against public evidence, and the Empirank audit found 479 of those against 24 contradicted claims.
Can a brand edit what ChatGPT says about it?
A brand cannot edit an AI answer directly. The record behind the answer is the part a brand controls: its own pages, its listings in registries and directories and the coverage published about it. Correcting a wrong claim means changing those sources and then re-running the same prompts to check the answer.
How often do AI answers get facts about a business wrong?
In the Empirank audit of one OpenAI configuration, 1.9% of testable claims were incorrect, and 11.0% of the recommended businesses carried at least one incorrect claim. A further 38.1% of claims could not be verified. The figures describe local businesses in August 2026 and may differ for other models and categories.
Why are smaller brands more exposed to wrong claims?
LLMs perform worse on less popular facts, according to the arXiv study by Mallen and colleagues, and larger models did little for long-tail knowledge. A smaller brand has fewer pages that state its facts, so a model has less to check against. The study tested models from 2022 and 2023, which makes it evidence for a mechanism and no measure of today's error rate.
How long does a correction take to show up in an AI answer?
The Empirank audit does not measure it, and its suggested retest after source corrections has no reported result behind it. Models refresh on their own schedules and third-party pages need time to be recrawled, so the practical route is a fixed re-run cycle on the same prompts, with each claim relabelled after every run.
Teams that want to see how AI assistants describe their brand today can start with a free AEO Monitoring scan from AI, TELL ME!.


