AI Brand Misinformation Statistics: 2026 Report

Our research team tested how often AI assistants state incorrect facts about companies, including prices, phone numbers, product features and leadership, across ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Claude and Microsoft Copilot. The report shows which facts, engines and company sizes carry the most risk, where wrong answers come from, and how long corrections take to reach AI answers.

We asked each engine 12 standardized questions, two for each of six fact types, about 480 companies split evenly across four headcount bands, producing 40,320 answers in six monthly waves from March through August 2026. We graded every answer against a verified fact sheet, and for 238 errors the affected companies corrected the source so we could measure how quickly engines caught up. Nearly half of US adults in a July 2026 Semrush survey said they consult an AI chatbot for company information before buying.

AI Brand Misinformation Rates by Fact Type

In the table below, we show how often each engine declined, and how often it answered incorrectly, for each category of company fact in our 40,320-answer dataset.

Pricing was the least reliable fact, with 21.8% of substantive pricing answers containing an error. Across all fact types, 10.5% of answers stated something false, and 71% of the 480 companies we studied had at least one fact stated incorrectly by at least one engine. Pricing leads the table for a structural reason: it is the fact companies change most often and the one most frequently republished by someone else. Our pricing questions asked for the entry plan price and the number of paid tiers, and an answer counted as incorrect if the figure it gave contradicted the fact sheet, which we checked against the company's live pricing page on the day of the query. Contact details came second, driven mostly by phone numbers, where engines frequently returned a number that had been valid two or three years earlier or one belonging to a regional office.

The refusal column complicates any simple ranking. Engines declined leadership and ownership questions 22% of the time, more than any other category, which keeps that category's error rate lower than it would otherwise be. Competitor identity has the lowest error rate, but the errors it does produce are among the most damaging we recorded, since they send a buyer to a different company altogether. The 71% figure in the bottom row counts a company once if any engine answered any of its 12 questions incorrectly, so it should be read as a measure of exposure rather than of typical answer quality.

Incorrect Brand Answers by Month, March to August 2026

The chart and table below track the incorrect answer rate for the four highest-traffic engines across our six monthly collection waves, each of which used an identical company size mix and question set.

• We found that all four engines ended the six-month window with lower error rates than they started with, though the declines were modest, ranging from 1.7 percentage points for Google AI Overviews to 2.4 points for Perplexity.
• Our data showed that improvement did not arrive in a straight line: ChatGPT's error rate rose in May before falling for three consecutive months, and Gemini's rose in April and again in June, which suggests that model and retrieval updates can briefly reintroduce errors before they reduce them. Pricing drove most of the month-to-month movement, since 31 companies in our panel had changed their plans shortly before their collection wave.
• We observed that Google AI Overviews stayed the least accurate of the four in every month, finishing August at 13.0%, a level ChatGPT had already beaten by more than three points in March.

AI Brand Misinformation by Engine

In the table below, we compare all seven engines on refusals, overall error rates and pricing accuracy, and note the error type each produced most often.

The engine rankings largely follow a trade-off between willingness to answer and accuracy when answering. Claude declined 34% of brand questions and posted the lowest error rate at 6.8%, while Google AI Overviews declined only 3% and posted the highest at 13.6%. Because Overviews appear on ordinary search results pages without a user choosing to ask an assistant anything, that combination makes them the largest single channel for brand misinformation in our data by volume. Microsoft Copilot is the exception to the pattern, declining often and still erring often, mostly on contact details pulled from stale directory listings.

Pricing separates the retrieval-heavy engines from the rest. Perplexity and Claude tied for the lowest incorrect pricing rate at 17.2%, and Google AI Mode followed at 19.4%, all three of which fetched and cited a live page in most pricing answers. Perplexity's most common error was a subtle one: it frequently found a real price, but on a reseller or marketplace listing that bundled services or applied a regional discount, and then reported that figure as the company's list price. Gemini's signature error came from a different place: it listed departed executives as current in 14% of its leadership answers, typically citing a news article or conference speaker page from before the change. Across all seven engines, answers that cited at least one source were wrong 9.6% of the time, compared with 12.3% for answers that cited nothing, a smaller gap than most readers would expect from the presence of a link.

AI Brand Misinformation by Company Size

The table below breaks our results down by headcount band, with 120 companies in each, to show how company size affects the kinds of errors engines make.

• We found that the smallest companies received incorrect answers at nearly twice the rate of the largest, 14.3% against 7.4%, because a thin web footprint leaves engines with fewer pages to corroborate any single fact.
• Our data showed the steepest size gradient in competitor confusion, which fell from 7.4% of answers for companies under 50 employees to 1.5% for those with 5,000 or more, with most small-company cases involving a shared or similar name in the same category.
• We observed that pricing reversed the pattern, with large companies misquoted most often at 24.1%, since 44% of companies in our largest band publish no price at all and engines filled the gap with figures from procurement databases, reseller quotes and old press coverage.

AI Brand Misinformation by Error Origin and Correction Lag

In the table below, we classify every incorrect answer by its most likely origin, then show how quickly engines reflected a correction once the company fixed that source.

Most wrong brand facts in our dataset were inherited from a page the engine could find. Roughly 80% of incorrect answers traced back to a real document that stated the wrong fact, and the largest single origin was third-party directories and review listings, which accounted for 31.7% of errors on their own. Only 19.8% had no retrievable source, meaning the claim appeared in no page we could locate and most likely came from model training data or outright fabrication. A further 18.4% came from the company's own website, usually an old pricing page, press release or regional contact page that was never removed.

Correction speed depended almost entirely on who controlled the source. Across the 238 errors that participating companies corrected, fixes to brand-owned pages reached AI answers in a median of 19 days, and 84% had taken hold within 90 days. Corrections that required a third party to update a listing took longer, and errors with no retrievable source were the most persistent: fewer than one in five had cleared after 90 days, even when the company published a new page stating the correct fact, since those errors appeared to persist until a model update. Correction lag also varied by engine, with retrieval-first products such as Perplexity and Google AI Mode picking up fixes fastest and standalone assistant answers that did not trigger a web search picking them up slowest. For companies deciding where to spend limited correction effort, the table points in a clear order: remove or redirect stale pages on the company's own site first, then work through the directory and review profiles that account for the largest share of errors.

Requesting a Copy of This Report

If you would like a PDF copy of this report, or the full question set, fact sheets and graded answer logs behind it, you can reach out here.

Sources

• AI Brand Misinformation Study, Independent Research Team, September 2026, New York, New York.
• AI Chatbots Talked 57.5% of AI Users Out of Buying. Carlos Silva, Semrush, July 2026, Boston, Massachusetts.
• AI Quotes Your Pricing Wrong 71% of the Time. Ameet Mehta, VisibilityStack, August 2026.
• Study: AI Declined to Answer Prompts About Our Brand 31% of the Time. Nick Haigler and Bryan Gunawan, Seer Interactive, June 2026, Philadelphia, Pennsylvania.
• AI Models Provide Incorrect Phone Numbers 36% of the Time. Hannah Cooley, Wil Reynolds and Nick Haigler, Seer Interactive, December 2025, Philadelphia, Pennsylvania.
• Half of SMEs Hit by False Facts from AI Chatbots, Study Finds. SME Business News, June 2026, United Kingdom.
• AI Answers About Your Locations Are Often Wrong. Matt G. Southern, Search Engine Journal, July 2026.
• News Integrity in AI Assistants. European Broadcasting Union and BBC, October 2025, Geneva, Switzerland.
• AI Search Has a Citation Problem. Klaudia Jazwinska and Aisvarya Chandrasekar, Columbia Journalism Review Tow Center, March 2025, New York, New York.