Copy for Originals · Guide · AI search and reputation
This page sets out what was measured in 2026 about AI answers on companies, what the platforms offer and where the courts stand, with every number checked against its source on October 24, 2026. It ends with a nine-step audit a company can run.
AI assistants do get facts about companies wrong, and how often depends on what is asked. Two 2026 studies, both by firms that sell related services, came out far apart. Seer Interactive, an agency, audited answers about its own brand on six platforms: 2.8% were wrong and 31.2% were non-answers. Reputation Age and Bubbling found 40% of the half-year figures ChatGPT and Gemini gave for 39 of France’s largest listed companies (the CAC 40) wrong, mostly outdated or invented. We found no platform channel built for companies to correct an answer, and none that promises a fix; the closest is Perplexity’s form for reporting inaccurate answers. What a company can do: test with fixed prompts run several times, trace each error to its source, fix the pages and listings it controls, report the answer, and retest. Full Fact, tracking civic facts, saw a corrected figure largely disappear from answers in one case, while in another some models still gave a figure replaced more than a year earlier. We found no measurement of how long a fix takes for company facts. Courts in the US and Germany have begun to rule, mostly at early stages.
In the two 2026 measurements of company facts we found, the error rate ranged from 2.8% of answers about one agency’s own brand to 40% of financial figures about France’s largest listed companies. Both come from firms that sell services related to the subject, and we found no independent replication.
The self-audit. Seer Interactive tracked 1,562 questions about its own brand, collected daily from May 1 to June 9, 2026, on Google AI Mode, Google AI Overviews, ChatGPT, Claude, Gemini and Perplexity, and scored the 28,123 responses against 87 facts about Seer. On average, 2.8% of answers were wrong and 31.2% were non-answers; when a platform did answer, it was right 95.9% of the time.
| Platform | Wrong | No answer | Right, when it answered |
|---|---|---|---|
| Google AI Overviews | 5.7% | 28% | 92.1% |
| Gemini | 4.3% | 28.2% | 94.1% |
| Perplexity | 3.3% | 29.2% | 95.3% |
| Google AI Mode | 1.7% | 24.1% | 97.8% |
| ChatGPT | 1.3% | 27.1% | 98.2% |
| Claude | 1.1% | 46.3% | 98% |
| Average | 2.8% | 31.2% | 95.9% |
Three limits apply. It is one company, measured by itself. A partial mention counts as accurate: if an answer names Philadelphia as Seer’s headquarters inside a longer answer, that attribute is marked right. And Seer does not say whether it collected answers through the platforms’ APIs or their consumer apps, who graded them, or which location the questions came from; the data came from Seer’s Scrunch data set.
The CAC 40 study. Reputation Age, a communications agency, and Bubbling, which describes itself as measuring how brands are cited in AI answers, used an automated persona of an economic journalist to ask ChatGPT and Gemini for the first-half 2026 results of 39 CAC 40 companies: 8,671 conversations and 34,684 responses between August 25 and 28, 2026. The study found that “40% des chiffres générés par ChatGPT et Gemini sont incorrects” (40% of the figures generated were incorrect). The share of correct answers ran from 67% for net income to 42% for published growth, and by company from 81% for Capgemini to 32% for Crédit Agricole. ActuIA, which reported the study, noted that “The protocol is only partially published and the two authors sell services related to the subject” and that the model versions and use of web search are not stated.
Other tests, not about companies. In a European Broadcasting Union and BBC study of more than 3,000 news answers from ChatGPT, Copilot, Gemini and Perplexity, published in October 2025, “20% contained major accuracy issues, including hallucinated details and outdated information”. On the other side, Google said in May 2024 that AI Overviews’ accuracy was on par with featured snippets, without giving a figure.
What it means. There is no single error rate to quote. A simple fact about one firm and a specific figure from a half-year earnings release are different questions, and they produced very different results: when Seer’s platforms answered, about 4% of answers were wrong (our calculation from Seer’s 95.9%), against 40% of the CAC 40 figures. The only number that tells you about your company is one you measure yourself, with the method in the steps below.
Outdated facts, services the company does not offer, wrong numbers, mix-ups with another company of the same name, and claims that no cited source supports. How these types split overall is not known, because no 2026 study we found separates a stale fact from an invented one.
In Seer’s audit, the errors Seer gives counts for were its former office in San Diego (223 times), services Seer does not offer such as reputation management (109 times) and a wrong headcount (85 times), which Seer files under “Fix First”, and an incorrect director’s title (3 times), which it calls housekeeping. A narrative about “High account turnover” was built from a single client review posted in 2018. The San Diego error was out of date: Seer describes it as AI “saying Seer’s San Diego office is still open”. Seer does not say whether the services and headcount errors were out of date or invented.
In the CAC 40 study, the wrong figures fell into two groups. On every one of the seven financial indicators, “outdated or invented” figures were more common than real figures that answered a different question:
| Indicator | Correct | Another official figure, wrong for the question | Outdated or invented |
|---|---|---|---|
| Net income | 67% | 13% | 19% |
| Net debt | 66% | 8% | 26% |
| Revenue | 65% | 6% | 29% |
| Organic growth | 64% | 12% | 25% |
| Free cash flow | 62% | 8% | 29% |
| EBITDA | 57% | 7% | 36% |
| Published growth | 42% | 24% | 34% |
The second kind of error, the right document but the wrong number, was a recurring pattern for three companies: 46% of checks for BNP Paribas, 43% for Crédit Agricole and 26% for EssilorLuxottica. The study’s own web page says the models “parfois” (sometimes) find the right document and pick the wrong indicator.
Mix-ups and unsupported claims appear in news reports and court cases. CBC News reported in July 2026 that Google’s AI Overview had said a bakery in Millbank, Ontario was closing; “the AI had mistaken Anna Mae’s for a bakery of the same name in Illinois”. In a case decided by the Munich Regional Court in May 2026, the legal news site LTO reported the court’s view that some of the AI’s claims did not appear in the linked sources at all. In Keene v. Google, the AI Overviews cited Wikipedia, and the court wrote that “Only after a user read the linked Wikipedia article would any contradiction become apparent”.
These cases were reported because they went wrong, so they say nothing about frequency, and the studies can’t settle it: Seer’s wrong answers cover “a stale value pulled from an old source or something the model invented with no basis”, and the CAC 40 study’s largest group is “donnée obsolète ou inventée” (outdated or invented data).
What it means. Expect two kinds of error that need different handling. Errors with a source, such as an old office, an old figure or one old review, can be traced and the source fixed. Errors without one, such as a mix-up with a namesake or a claim the cited page doesn’t make, can only be reported and outweighed with clearer, more consistent facts elsewhere. Your audit should record which kind each error is, because the studies don’t.
The company studies don’t say why. Research on how models learn facts suggests one likely reason: models answer less reliably about things that are rarely written about, and applying that to companies is our inference. What Seer’s audit does show is that declining to answer was far more common than answering wrongly: 31.2% non-answers against 2.8% wrong answers. Seer counts as a non-answer any response where “The model declined to answer or told the user to check with the brand directly”.
The rate varied by platform and by question. Claude gave no answer 46.3% of the time and Google AI Mode 24.1%. Direct questions about Seer were answered accurately 90.4% of the time and indirect ones 88.7%. Comparative questions (“How does Seer Interactive compare to other marketing agencies?”) were accurate only 18.8% of the time, with no answer 80.3% of the time.
The research is about facts in general. Kandpal and colleagues (ICML 2023) found that a model’s ability to answer a factual question “relates to how many documents associated with that question were seen during pre-training”, and Mallen and colleagues (ACL 2023) that models “struggle with less popular factual knowledge”. Neither looked at companies. Our inference: a rarely written-about firm is less likely to be known from training, and more dependent on what an assistant can find and read when it answers.
What it means. For many companies the bigger problem is silence, not error, and it is worst on comparison questions, which are likely the ones buyers ask when choosing (our view). A non-answer is not a correction problem; it is the visibility problem covered in What decides whether an AI recommends you. Record non-answers separately from wrong answers in your audit, so you fix the right one.
Seer groups the causes of wrong answers into three: a stale page the assistant read while answering, what the model learned in training, and a misreading of the company’s own site. Some answers make claims that the sources they cite don’t contain.
Seer attaches an estimated time to fix each. Training data has “No citation or source to trace” and takes months, tied to the next retraining. Live retrieval from a stale source is “Traceable and fixable” in days to weeks once the source is updated. A misread of your own site is, in Seer’s estimate, among the fastest to fix, by adding “more explicit language to existing pages”. These times are estimates, not measurements.
In both studies, the companies’ own websites were the most-cited sources. In Seer’s audit, seerinteractive.com had over 35% of 219,349 citations across 9,240 domains. In the CAC 40 study, 78% of the citations, nearly all of them ChatGPT’s, pointed to the companies’ own sites, and 0.08% to press articles. Emmanuel Dollé, co-founder of Bubbling, offered an opinion, not a measured result, in an op-ed by Reputation Age published on Boursorama: the assistants “retiennent souvent le chiffre le plus visible, généralement celui de la première page” (often keep the most visible figure, usually the one on the first page) and restate it out of context.
In one test, one page was enough. Lily Ray, an SEO practitioner, published an invented claim on her own site and wrote in April 2026 that “Within 24 hours, Google’s AI Overviews was confidently serving this fabricated information back to users”; hers was the only source making the claim.
Tracing has limits. A 2023 study of four generative search engines, older than today’s, found that “a mere 51.5% of generated sentences are fully supported by citations and only 74.5% of citations support their associated sentence”. The court cases above show the same gap, and Wolf River Electric, a business suing Google, alleges that “none of the referenced materials in fact contained the information Google claimed they did”.
What it means. Start with your own site, the most-cited source in both studies. State the facts that matter as plain sentences on pages a crawler can read: where you are and aren’t, what you offer and don’t, how many people work there, and current figures near the top. Then check the old places your name appears: former office listings, old press releases, a lone review. How to write content that LLMs retrieve and cite covers how to write those pages, and Can AI read your website?, our free tool, checks whether the crawlers can reach them.
Not necessarily. Full Fact, a fact-checking non-profit, found that “models sometimes gave different answers to the same question on different days”, so a single wrong answer is one observation, not a rate. How much brand recommendations vary is covered in What decides whether an AI recommends you.
What it means. One screenshot of a wrong answer tells you the error exists, not how often people see it. No study we found has tested how many runs a fact audit needs. For brand-visibility monitoring, the University of St. Gallen recommends “at least 7 runs per prompt per day”. A one-off fact audit is a different job, and our own suggestion is lighter: at least seven runs of each prompt on each platform in total, spread over several days, with the raw answers and dates saved. Treat that as a floor, not a standard.
Not as a company, in the help pages and policies we read. Every channel we found is either a feedback route that promises no correction (Perplexity’s report form for inaccurate answers is the most specific), or a privacy or legal route for personal data or illegal content, where the platform promises only to “consider” or make a “reasonable effort”. That is what the pages we read show as of October 24, 2026, not proof that no such channel exists.
| Platform | For anyone: feedback | For named people and legal issues | What the platform promises |
|---|---|---|---|
| ChatGPT (OpenAI) | Thumbs down below a response; “Report” flags possible breaches of terms or law. The Report Content form’s reasons do not include inaccurate information. | Correction or removal of “factually inaccurate information about you” via privacy.openai.com or dsar@openai.com; a “Remove my personal data from ChatGPT responses” request. | “we will consider your request based on applicable law and the technical capabilities of our models”. On approval: “Not necessarily”. |
| Google AI Overviews | “Report a problem”. | Report Content for Legal Reasons; the page does not mention AI features. | “Your feedback helps us improve AI Overviews for everyone”. For legal notices: “consider blocking, limiting, or removing access”. |
| Google AI Mode | Thumbs up or down. | As above. | “so we can continuously improve”. |
| Gemini | — | Under laws such as the GDPR, ask for “inaccurate personal data in Gemini Apps’ responses to be corrected”; in Gemini, More, then Report legal issue. | No correction promise on the page. |
| Microsoft Copilot | “use the Report or Feedback features in Copilot”. | “Report a Concern” page for legal concerns. | No correction promise on the page. |
| Perplexity | Flag icon below the answer, or a support ticket; issue types include “Misinformation” and “Outdated information”; three-dot menu, “Report”; support@perplexity.ai for other issues. | — | “happy to take your report to improve the quality of your responses”. |
| Claude (Anthropic) | Thumbs down, or the feedback email. | A correction request for “factually inaccurate personal data relating to you”. | “a reasonable effort to correct this information—but ... it may not always be possible”. |
Table notes: — means the pages we read on October 24, 2026 do not describe this route. For Microsoft 365 Copilot Chat (work and school accounts), Microsoft says thumbs-up and thumbs-down feedback is used “to improve Copilot”, but “We don’t use this feedback to train the foundation models used by Copilot”.
Perplexity’s report form, which asks for a “Description of the error and expected result”, is the closest thing to a correction request we found, though it promises only to “improve the quality of your responses”.
Two related routes change sources, not answers. In Google Business Profile, owners can edit their own profile, and Google says such changes “usually take about 10 minutes to show up” and can take up to 30 days. Anyone can suggest an edit to a profile they don’t manage; the help page gives no timing for that. Neither help page says an edit changes what AI Overviews or AI Mode say. IndexNow tells Bing and some other search engines that a page has changed; nothing on its site says any AI assistant will refresh an answer after a ping.
Google, OpenAI, Microsoft and Anthropic also disclaim accuracy in the pages we read. Google writes that “AI Overviews can and will make mistakes”, that AI Mode “doesn’t always get it right”, and, in its Gemini privacy hub, that LLM experiences, Gemini Apps included, “can hallucinate”. OpenAI, Microsoft and Anthropic say the same.
What it means. Use the feedback buttons anyway, with the correct fact, because they are the only direct route and cost a minute. But plan as if they do nothing, since no source we found shows that they lead to a correction. If the wrong answer is about a named person, such as a founder, a partner or a doctor, that person has a stronger route: the privacy request. Save a screenshot, the date and the prompt each time you report.
Sometimes. There is evidence both ways, and we found no published measurement of how long a correction takes for company facts.
Seer writes that “Once you update the source of outdated information, you should see AI model responses reflect those changes within days”, and that the whole process “will take weeks to months”. That is a forecast. Seer says in the same post that it had earlier succeeded in “changing the narrative in AI” about “High account turnover”, but this study reports no measured retest after fixes.
Full Fact has the only independent evidence we found, from civic and news facts, collected by querying the models through their developer APIs rather than the consumer apps. In routine checks of claims it was fact-checking, begun in February 2026, Full Fact found that “In some cases the LLMs corrected themselves days later”, but in “several instances” chatbots “continued to respond with incorrect information”. After it published fact checks on a fake Ryanair picture and a miscaptioned Green Party video, “All the LLMs which initially made errors corrected themselves”. During Full Fact’s benchmark monitoring, the House of Commons Library published a wrong child poverty figure, and “When the Library corrected the error, the incorrect statistic largely disappeared from model responses”; Full Fact wrote that this “illustrates how quickly models can learn and adjust their answers when source information changes”, without giving a timing. It adds a caveat: some answers repeated the wrong figure without citing the Library, so the error may have spread through other sites, and “If so, correcting the original does not necessarily clear the record”.
When the UK Prime Minister changed on July 20, 2026, “Most models updated promptly, but DeepSeek was still identifying Sir Keir as both Prime Minister and Labour leader as late as September”. And “several chatbots cited a weekly rate of £8.86, which was superseded by £9.95 in June 2025”: “Gemini 3.1 Pro and Claude gave the outdated figure outright in some responses” in data collected from July 15 to September 22, 2026, more than a year after the change.
What it means. Fix the sources anyway. It is the only step with any evidence of working, and the only one fully in your control. Then retest with the same prompts after two to four weeks, and again after a model update, rather than assuming the fix took. If an error survives a corrected source, look for a second source repeating the old fact, or assume it comes from training data, which in Seer’s estimate takes months to change.
Not much yet, and it depends on whether the false statement is about the company or about a named person. This is a summary of public rulings and complaints as of October 24, 2026, not legal advice.
Data protection covers people, not companies. The GDPR “does not cover the processing of personal data which concerns legal persons and in particular undertakings established as legal persons” (Recital 14). A named person does have “the right to obtain from the controller without undue delay the rectification of inaccurate personal data” (Article 16). noyb’s complaints against OpenAI, filed in Austria in April 2024 and Norway in March 2025, are listed as pending on its case pages, with the latest entries in 2025. In the Norwegian case, the false answer stopped after OpenAI added web search, a platform change rather than a source fix, and noyb warned that “the incorrect data may still remain part of the LLM’s dataset”.
United States. OpenAI won summary judgment in Walters v. OpenAI, a libel suit in Georgia, on May 19, 2025. In Keene v. Google, where one AI Overview said the plaintiff was “serving a life sentence without parole for the murders of three women”, a federal court in Illinois denied Google’s motion to dismiss as to two of four overviews on September 14, 2026. In Starbuck v. Google, over statements he says Google’s AI tools, including Bard and Gemma, made about him, a Delaware court denied Google’s motion to dismiss “in its entirety” on July 24, 2026; Fox News reported that the case advances to discovery. Wolf River Electric alleges that Google’s AI defamed it; in January 2026 a federal court sent its case back to state court because Google filed its removal too late, without ruling on the merits.
Germany. On May 28, 2026, the Munich Regional Court granted a Munich publishing house and its subsidiary an interim injunction against Google over an AI Overview, on the basis of the company’s right to its reputation. The court held that the AI Overview contains Google’s own statements. Google said in June 2026 that the decision “is not yet final” and that it planned to appeal, calling the case one that “focuses on specific and narrow errors”. On September 24, 2026, the Aachen Regional Court, citing Munich, issued an interim order over AI Mode attributing another person’s history to a doctor. According to a lawyer’s blog, the court based Google’s responsibility on providing AI search and the answers it generated, and the doctor’s unsuccessful complaints added to it. A third ruling, by the Hamm Higher Regional Court on May 12, 2026, concerns a company’s own website chatbot, which falsely told users the two doctors behind the company were specialists in plastic and aesthetic surgery. The court treated those answers as the company’s own misleading commercial practice. The court granted leave to appeal to the Federal Court of Justice, and CMS’s case tracker listed the ruling as not final as of June 2026.
What it means. In the cases so far, companies have relied on defamation law in the US (Wolf River) and on a company’s right to its reputation in Germany (Munich). Whether either route exists where you are is a question for a lawyer. A named founder, partner or doctor also has the GDPR route, though noyb’s complaints are still pending. In Walters, not asking OpenAI for a correction or retraction meant no punitive damages under Georgia law. In Keene, an earlier report through Google’s “report a legal removal issue” form helped the claim survive a motion to dismiss: the court held it plausibly supported knowledge or reckless disregard of the falsity. So report serious errors through the platform’s legal form, keep the record, then talk to a lawyer. And check what your own site’s chatbot says: in Hamm, the court held the company responsible for it.
The protocol below follows Seer’s: “identify your brand’s critical attributes, build test prompts against them, and track accuracy daily before chasing anything else”. We added repeated runs, source tracing and a retest. Even a small audit is a lot of answers: 20 facts in three prompt shapes, run seven times on seven platforms, is 2,940 answers to save and score (our arithmetic). Start smaller if you need to (see step 3).
It depends on the question. In Seer Interactive’s 2026 audit of answers about its own brand, 2.8% of 28,123 answers across six platforms were wrong and 31.2% were non-answers; Google AI Overviews was wrong 5.7% of the time and ChatGPT 1.3%. In a study by Reputation Age and Bubbling of 39 of France’s largest listed companies (the CAC 40), 40% of the half-year financial figures ChatGPT and Gemini returned were wrong. Both studies come from firms that sell related services.
Not through a dedicated channel that we found. As of October 2026, ChatGPT, Google AI Overviews, AI Mode, Copilot, Perplexity and Claude offer feedback buttons, and OpenAI, Google and Anthropic offer privacy routes for inaccurate personal data about individuals. In the help pages and policies we read, none is a correction route for companies, and none promises a fix. Use the feedback buttons, fix the sources the answer relies on, and retest.
No. Recital 14 says the GDPR does not cover data that concerns legal persons such as companies. A named person, such as a founder, partner or doctor, does have a right under Article 16 to rectification of inaccurate personal data, and OpenAI, Google and Anthropic accept such requests, but in the policies we read none promises to make the correction. noyb’s complaints against OpenAI over this were still listed as pending in October 2026.
We found no published measurement for company facts. Seer Interactive forecasts days to weeks for errors that come from a source an assistant reads live, and months for errors learned in training, but those are estimates. Full Fact, testing news and civic facts, saw one corrected statistic largely disappear from answers, and an outdated rate still stated outright by some models more than a year after it had been superseded.
Cases are under way, and most rulings so far are at an early or interim stage. The company cases are in Germany, where the Munich Regional Court granted a publishing house an interim injunction against Google over an AI Overview in May 2026, which Google said in June 2026 it would appeal, and in the US, where Wolf River Electric’s defamation claim against Google was sent back to state court on procedure in January 2026. Individuals’ cases show how courts treat AI answers: OpenAI won summary judgment in Walters v. OpenAI in 2025, while Starbuck v. Google survived a motion to dismiss in full and Keene v. Google in part in 2026. In Walters, not asking OpenAI for a correction or retraction meant no punitive damages under Georgia law; in Keene, an earlier report through Google’s legal removal form helped the claim survive a motion to dismiss.
Non-answers are more common than wrong answers. In Seer Interactive’s audit, 31.2% of answers about its brand were non-answers, against 2.8% wrong, and comparative questions went unanswered 80.3% of the time. Research on language models finds they struggle with less popular facts. Applying that research to companies is our inference: a less-known company likely depends on pages an assistant can find and read when it answers.
All accessed October 24, 2026. Sources from companies that sell AI visibility, reputation or listing services are marked (vendor).
How to cite: Copy for Originals, “When ChatGPT or Google’s AI gets your company wrong: what the 2026 evidence shows and what you can do,” October 24, 2026, https://tools.copyfororiginals.com/when-ai-gets-your-company-wrong/.