An AI Is Wrong About Your Brand: How to Detect and Fix the Error
ChatGPT, Gemini or Perplexity getting your company wrong? Where the error comes from, how to fix it source by source, and realistic timelines.
Published 3 November 2026 · Sapian Metrics · This article will be updated as AI engines change their reporting channels.
A prospect emails you to say ChatGPT claims your company has shut down. Gemini credits your product to a competitor. Perplexity quotes a price you have never charged. You are not alone: according to the EBU/BBC study of October 2025, 45% of AI assistant answers contain at least one significant issue, and 20% a major factual error. The bad news: nobody can "edit" a language model's answer. The good news: in the vast majority of cases the error comes from a page the model read, and that page can be fixed. You don't correct an AI, you correct what it reads.
This article gives a four-step method: measure the extent, diagnose the origin, fix source by source, then track the correction over time with realistic timelines.
How common are errors about brands?
Short answer: errors are the norm, not the exception, but they cluster around a few topics (pricing first) and a few sources. According to Profound, the inaccuracy rate of claims about brands ranges from 4.3% to 6.8% depending on the engine, and 54% of brands have at least one false claim that cites their own website.
Three studies frame the issue without dramatising or minimising it.
The News Integrity in AI Assistants study, coordinated by the European Broadcasting Union and led by the BBC, had more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity evaluated by 22 public service media organisations in 14 languages. The result: 45% of answers had at least one significant issue, 31% a serious sourcing problem, 20% a major accuracy error. Gemini performed worst with 76% of problematic answers, mostly because of sourcing. The study covers news, not brands, but the mechanisms are the same.
On brands specifically, the analysis Where do inaccurate AI claims come from, published by Profound in August 2026 on more than 158,000 fact-checked claims, gives four useful figures. The inaccuracy rate per engine ranges from 4.3% to 6.8%. Pricing and billing make up 12% of evaluated claims but 24% of inaccurate ones. 54% of brands have at least one inaccurate claim citing their own content, and 51% at least one citing a third-party publication. Finally, ChatGPT and Claude share only 12.5% of their claims: an error on one engine does not predict the same error elsewhere, which means checking engine by engine.
Why do models make things up rather than admit they don't know? OpenAI explained it in Why language models hallucinate in September 2025: evaluation methods reward a guessed answer the same as a correct one and give no credit for "I don't know". Rare, arbitrary facts (a founding date, an executive's name, a price) are exactly the ones a model cannot infer from a statistical pattern. Unless your brand is very famous, your brand is a rare fact.
Diagnose: where does the error come from?
Short answer: an AI error has three possible origins, and each is fixed differently. Either the model read a wrong or outdated page (the error is sourced), or it answers from training memory (no source, stable answer), or it confuses your brand with another entity. The diagnosis takes an hour and determines everything that follows.
Here is the protocol we apply in our audits. It can be reproduced without any tool.
Ask the same question, phrased three different ways, on at least four engines (ChatGPT, Gemini, Perplexity, Google AI Overviews), with web search enabled where the option exists. Repeat each question three times in separate sessions: what we observe on our daily scans is that the same question does not always produce the same answer, and an error that appears one time in three is not handled like a systematic one. For each wrong answer, note the engine, the date, the exact wording, and above all the sources cited.
Then classify.
Case A, the error is sourced. The engine cites a page, and the error is on that page. This is the most frequent case and the most encouraging one, because it is actionable. The most common finding in our audits matches Profound's figures: the faulty page is often one of the brand's own pages, forgotten for two years, with an old price, a former executive or an old address, which the engine preferred over the current page because it answered the question more directly.
Case B, the error is from memory. No source cited, the answer is identical from one session to the next and persists even in search mode. The model is reproducing what it learned in training, from pages that may have disappeared since. This is the slowest case to fix.
Case C, the error is an entity confusion. The engine mixes your brand up with a namesake, a subsidiary, a former name or a competitor with a similar name. The signs: products you don't sell, a country you are not in, a foreign founding date. The remedy is explicit disambiguation, not a factual correction.
For more on measurement itself, our guide to measuring AI visibility explains how to build a stable question panel.
Fix it, source by source
Short answer: for a sourced error, fix the page, whether it is yours or a third party's, then force re-indexing. For a memory error, publish a page that answers the false question exactly, so the engine reads it first when it searches. For a confusion, publish a canonical facts page that sets you apart unambiguously.
When the faulty source is your own site
Correct the fact on the cited page. If the page is outdated, don't let it live on: update it or redirect it (301) to the current page. Check that visible dates and date tags are consistent, engines use them to arbitrate between two versions. Create or consolidate a canonical facts page (an About page or press kit) with the data AIs look for most often: legal name, founding date, executives, locations, product range, prices or price ranges, key figures. Align structured data (Organization, Product, Offer) with the visible text; Google explicitly recommends this consistency for its AI features. Then request re-indexing of the corrected URLs through Search Console's URL inspection and its Bing Webmaster Tools equivalent, which feeds Copilot.
An up-to-date llms.txt file helps steer agents reading your site towards the right pages; we have detailed its benefits and limits. Transparency note: the Sapian Metrics platform generates this file automatically, but you can write it by hand in an hour.
When the faulty source is a third party
Profound observes that errors from third-party media are concentrated: for a given brand, they generally come from about four sites, and the top 100 domains account for 40% of errors of this type. In other words, you don't have to fix the Internet, you have to fix four pages. Contact the publisher with the factual correction and a verifiable source; it is the most effective route. Prioritise the directories and reference databases engines cross-check: Wikipedia and Wikidata if you are listed, Google Business Profile and Bing Places listings (Microsoft states that Bing Places is the channel for correcting a business's information), and the comparison and review sites in your sector. AI-generated comparisons rely heavily on these third-party pages.
When the error is from memory
You cannot modify a model that has already been trained. You can, however, make sure that at answer time the engine prefers reading to remembering. Publish a page that answers word for word the question that triggers the error ("Has [Brand] closed?", "Who runs [Brand]?", "How much does [Product] cost?"), with the answer up top, dated and sourced. Multiply consistent third-party corroborations. And accept that the full correction will wait for the next model generation.
Official reporting channels: what actually exists
Here we need to separate the confirmed from the hoped-for.
Confirmed: Google offers a thumbs-down on every AI Overview followed by "Report a problem". Microsoft offers a reporting form for Bing results and points to Bing Places for business data, with no dedicated channel for Copilot answers. OpenAI offers a personal data removal form, which concerns natural persons, not companies, and aims at removal or filtering, not rectification: in the complaint filed by noyb with the Austrian authority in April 2024, OpenAI stated it could not correct false information, only block answers about a name.
Observed: Perplexity, Claude, Le Chat and Grok offer a thumbs-down or a report button in the interface. None of these engines documents a rectification channel for brands, and none commits to a timeline.
Unknown: the actual effect of an individual report on future answers. No engine publishes it. Report anyway, it costs thirty seconds, but don't build your correction plan on it.
When it is serious: defamation and legal action
Short answer: if the error accuses you of fraud, a conviction or a closure, document everything, formally notify the engine's publisher, and consult a lawyer. The law is unsettled: one company has had its lawsuit against Google move forward, one individual lost against OpenAI. We are not lawyers; what follows is a summary of public cases.
The most telling case for a company is Wolf River Electric, a solar installer in Minnesota. According to the summary published by law firm Quinn Emanuel, a Google AI Overview stated that the state Attorney General had sued the company for deceptive practices, which was false and which the cited sources did not say. Customers cancelled. The complaint, filed in March 2025, seeks more than 100 million dollars; in January 2026 the case was remanded to state court and is in the pre-trial phase.
Conversely, in Walters v. OpenAI, a radio host whom ChatGPT falsely accused of embezzlement lost in May 2025: the court held that no reasonable reader would have taken the answer as a statement of fact, and that the plaintiff showed no damages. In between, Starbuck v. Meta, where the plaintiff stressed he had notified Meta repeatedly without correction, ended in a settlement in August 2025.
Three practical lessons. Document: time-stamped screenshot, exact question, engine, mode (with or without search), sources cited, and for heavy cases a witness or a notarised record. Notify formally: the cases that move forward are those where the publisher was warned and did nothing. Quantify the harm: lost contracts, cancelled quotes, emails from worried customers. In Europe, the GDPR (Articles 5 and 16, accuracy and rectification) protects natural persons, executives included, but not the legal entity as such.
Realistic timelines and follow-up
Short answer: a correction on a page read on demand propagates within days to weeks, the time it takes for a new crawl. A training-memory error waits for the next model generation, several months. In both cases, you only know you have succeeded if you measure again, engine by engine.
Google states that its preview controls can take "days to months" to be reflected after a recrawl. That is the order of magnitude to keep in mind for any page correction, at Google as at the other engines that read the web on demand. Corrections on third parties add the publisher's own delay.
Follow-up is the part teams neglect. An error fixed on ChatGPT can survive on Gemini; the low overlap between engines measured by Profound almost guarantees it. That is why accuracy is one of the core GEO KPIs we recommend tracking, alongside citation rate. Transparency note: Sapian Metrics tracks answer accuracy across eight engines (ChatGPT, Claude, Perplexity, Gemini, Grok, Copilot, Le Chat, Google AI Overviews), question by question, from 55 € per month; a spreadsheet and an hour a week do the same job at small scale.
B2C and B2B: the errors are not the same
In B2C, the most frequent errors we come across concern opening hours, prices, product availability, a recall or a store closure, and they hit the purchase decision directly: our article on AI-readable product pages covers prevention on the catalogue side. In B2B and SaaS, they concern pricing (the top error theme at Profound, for 66% of brands with enough evaluated pricing claims), features, integrations, compliance and supposed acquisitions. A B2B buyer who reads "not GDPR compliant" or "was acquired by X" in an AI answer will not call you to check. The pricing page and the security page are therefore the first two to audit.
FAQ
ChatGPT says my company has closed, what should I do first? Check whether the answer cites a source. If so, fix or get that page fixed, then request re-indexing. If not, publish a dated page that explicitly states "[Brand] is in business", and strengthen third-party signals (Google Business Profile, directories, press).
Can I ask OpenAI or Google to correct a piece of information? Not in the sense of a guaranteed rectification. Google offers "Report a problem" on AI Overviews, OpenAI a personal data removal form reserved for natural persons. No engine commits to a correction or a timeline.
How long before the correction shows up? Days to weeks for an error coming from a page read on demand, once the page is fixed and recrawled. Several months for a training-memory error, the time for a new model version.
Why does the error show on one engine and not the others? Because each engine reads different sources: according to Profound, ChatGPT and Claude share only 12.5% of their claims about brands. You therefore need to diagnose and track engine by engine.
An AI is defaming my company, can I sue? Proceedings exist (Wolf River Electric v. Google is ongoing), but the law is not settled and one plaintiff lost against OpenAI for lack of demonstrated harm. Document, notify formally, quantify the harm and consult a lawyer.
How do I stop it happening again? Keep a canonical facts page up to date, remove or redirect outdated pages, align structured data with visible text, monitor the handful of third-party sites that feed answers about your brand, and measure again regularly. The five most common AI visibility mistakes overlap heavily with this list.
Conclusion
An AI that gets your brand wrong is neither a fatality nor a black box. In most cases the error has an address: a page, yours or a third party's, that the engine read. The work consists in finding it, fixing it, forcing a re-read, then checking on each engine that the correction has taken. If you want to know what the eight main engines say about your brand today, and where they get it wrong, our free audit gives you the picture in a few minutes.
Are you cited by AI?
Get a free audit of your visibility across ChatGPT, Perplexity and Google AI.
Start my free auditRelated articles
Reddit, Wikipedia, YouTube: Why AI Engines Cite These Sources and How to Be There Without Cheating
Reddit, Wikipedia and YouTube dominate the sources cited by ChatGPT, Perplexity and Google AI. The figures per engine, the reasons, and the method for being present cleanly.
Black Friday: How ChatGPT and AI Engines Pick the Deals They Recommend (2026)
15% of shoppers will hunt Black Friday deals through AI, and 44% of prompts verify brands and discounts. How the engines choose, and how to get chosen.
AI-Readable Product Pages: What Agents See (and Miss) on Your Site (2026 Guide)
Product pages average 66% machine readability, the lowest of all retail pages. The method to make yours extractable before Black Friday.
