The numbers
≈1 in 10
AI answers to a hard factual question are wrong: about 9% of just over 4,000 audited Google AI Overviews gave an incorrect answer
45%
of 3,000+ AI assistant answers about the news carried at least one significant issue; 20% had major accuracy problems such as hallucinated or outdated information
Source: EBU/BBC, News Integrity in AI Assistants, October 2025
39%
of audited Google AI Overviews were both correct and fully supported by their cited sources; only 67% of individual claims were grounded in the citations
34.8%
of answer variance comes from just re-asking the same question; one-off checks are statistical noise
Source: arXiv 2607.13304, 12,933-response study on brand-answer reliability
No study has published a direct answer to "what percent of small businesses are misrepresented by AI right now". The figure this site uses, roughly 1 in 10, is the Oumi audit's roughly 9% wrong-answer rate on just over 4,000 factual queries: a per-answer rate, not a per-business one, with the EBU/BBC audit's 20% major-accuracy rate on news questions as the upper bracket. PositionBird's monitoring fleet is collecting this data; we will publish it when the sample is large enough.
When it goes wrong, it goes to court
In 2025 a Minnesota solar contractor, Wolf River Electric, sued Google after an AI Overview falsely told searchers the state attorney general had sued the company; the business claims $110–210M in lost contracts, in one of several AI defamation suits filed in two years. In July 2026 an independent bookshop discovered, from a customer, that AI engines were describing it as "permanently closed" at a wrong address; the answer stood for weeks. In 2026 the Munich Regional Court (Landgericht München I, in Bavaria) ruled that AI Overview statements are Google's own statements, making Google liable for false ones, and a US federal court has ordered AI chat logs preserved as evidence in litigation.
The pattern in every incident: the business found out late, from a customer, and had no record of what the AI had been saying or for how long. The first US AI defamation case (Walters v. OpenAI) was dismissed in 2025. Winning in court is hard, which makes early detection and documentation, not litigation, the practical defense.
What can actually be fixed
"Getting a wrong AI answer corrected" means changing what the engine reads, then waiting for it to re-read. How possible that is, and how fast, differs sharply by engine:
| ENGINE | FIXABLE? | MECHANISM AND TIMESCALE |
|---|---|---|
| Perplexity | Yes (partial) | Cites exact source pages; owned-page fixes re-crawled in 24–72h, third-party sources days–2 weeks |
| Google AI Overviews | Partial | Source and index dependent, typically weeks; the feedback button rarely resolves a claim on its own |
| ChatGPT (browsing) | Partial | Days–weeks when it retrieves your corrected sources |
| ChatGPT (no browsing) | No | Answers from training data; changes at the next retrain, months out |
| Gemini, Claude | No | Answers from training data; fix sources for the long term and document the wrong answers meanwhile |
Two things that don't work, despite common advice: llms.txt (Google's documentation confirms it has no effect on Search or AI Overviews, and major crawlers largely ignore it) and the answers' report/feedback buttons (feedback, not overrides; one documented case saw a formal cease-and-desist ignored until a court intervened).
How we measure it
Because a third of answer variance comes from resampling alone, a single check proves nothing. PositionBird's methodology: each business fact (status, hours, prices, policies) is checked with 3 differently-phrased questions × 3 samples per engine per weekly cycle; an LLM judge classifies each answer against the merchant's verified fact sheet; a fact is flagged only when the Wilson lower bound of its contradiction rate clears a threshold; and a claim is counted "misrepresented" only after two consecutive flagged cycles. Every sampled answer is stored with a SHA-256 hash and timestamp, so each data point is independently re-verifiable. The sources engines draw those answers from are measured the same way; see which sources AI engines cite for shopping questions, with the raw data.
Our own fleet measurement, the direct "% of businesses misrepresented right now" number nobody has published, is being collected across monitored stores and will be published here once the sample is large enough to be defensible. Monitoring your store adds it to the measurement.
Sources
- News Integrity in AI Assistants. European Broadcasting Union with the BBC, 21 October 2025. Journalists from 22 public-service media organisations in 18 countries evaluated more than 3,000 answers from ChatGPT, Copilot, Gemini, and Perplexity.
- Oumi's Study Finds 50% of AI Overviews Untrustworthy. Oumi, 14 April 2026. SimpleQA-based audit of just over 4,000 Google AI Overviews, checking both the answer and whether each claim is supported by the cited sources.
- Where Does the Noise Come From? A Variance-Components Decomposition of Non-Determinism in LLM Brand Answers. Dmitrij Żatuchin, arXiv:2607.13304, 2026. 12,933 responses across 20 brands, 8 languages, and 3 models, decomposed into resampling, paraphrase, model, and language variance.
- Minnesota Solar Company Sues Google Over AI Summary. Government Technology, 13 June 2025 (Wolf River Electric v. Google). An AI Overview falsely said the Minnesota attorney general had sued the company; the complaint claims $110–210 million in lost business.
- Landmark German ruling declares Google AI Overviews are Google's own words and makes it liable for false answers. The Decoder, 2026 (Landgericht München I, Bavaria). The Munich Regional Court held that AI Overviews are Google's own statements rather than third-party content, so Google is liable for false ones.
Is AI wrong about your business right now?
Check one fact free in 30 seconds, or monitor everything weekly with alerts and a court-ready evidence vault.
Dealing with an incident right now? Step-by-step guides: