# When AI Cites You Wrong: The Underestimated Reputation Risk > AI search engines get sources wrong in over 60% of cases — and sound confident. Why misquotes are a bigger risk for you than invisibility. Most businesses worry about being *overlooked* by AI. The bigger risk is subtler: that AI gets you *wrong* — and sounds completely convincing while doing it. Wrong opening hours, a service you don't even offer, an outdated price, a mix-up with a competitor. The customer reads it, believes it and never gets in touch. Or they get in touch with the wrong expectations. ## How often AI gets it wrong The error rate is high and well documented: in March 2025, Klaudia Jaźwińska and Aisvarya Chandrasekar of the Tow Center for Digital Journalism at Columbia University published the study *"AI Search Has a Citation Problem"*. They tested **eight leading AI search tools** — including ChatGPT Search, Perplexity, Gemini, Copilot and Grok — with **1,600 standardized queries**. The result: in **over 60% of cases** the AI search engines cited the source incorrectly. The spread is dramatic. Perplexity performed best, with an error rate of 37%. Grok-3, the paid version, came in at **94% wrong answers**. ## Made-up links included Particularly tricky: the models don't just invent facts, they invent sources too. In the tests, Grok-3 sent users **154 times to a dead 404 error page** — more fabricated, broken links than correct citations. A link that looks convincing but leads nowhere isn't a gain for your business, it's a loss of trust. ## The real problem: blind confidence The most dangerous finding of the study isn't the error rate itself, but *how* the errors are presented. The researchers call it "blind confidence": instead of admitting a knowledge gap, the models phrase things with extreme self-assurance — and are factually wrong. Out of 134 wrong citations, ChatGPT used a hedging phrase like "it appears" or "possibly" only **15 times**. In all other cases, the wrong answer sounded just as assured as a correct one. Only Microsoft's Copilot was an exception, preferring to decline an answer when the facts were thin rather than guess. For the person in front of the screen, this is fatal: when the machine answers in a fluent, self-assured tone, there's no warning signal. Right and wrong look identical. ## Why this hits your business in particular At their core, AI models are probability machines: they calculate the most plausible next statement. When such a model encounters patchy, contradictory or unstructured information about your business, it fills the gaps — not with truth, but with what's statistically nearest. If your address appears differently on three platforms, if your range of services is nowhere clearly declared, if your details contradict each other: that's exactly when the machine starts to guess. The consequence is uncomfortable, but unambiguous: **the burden of proof is on you.** It's not enough to have correct information somewhere. It has to be available so unambiguously, free of contradiction and machine-readable that AI has no room left to guess. ## Clarity is the defense The antidote is called [Entity Clarity](https://btlabs.dev/en/posts/dual-channel-visibility): your brand, your services, your data must appear across the entire web as *one* coherent, contradiction-free picture. A single, well-maintained data source, prepared in a structured way (for instance via JSON-LD), delivered consistently — that removes the ambiguity that otherwise trips AI up. How this machine-readable foundation works technically, we describe [in detail elsewhere](https://btlabs.dev/en/posts/from-seo-to-geo-architecture). This is exactly where [**btlabs Core**](https://btlabs.dev/en/ai-website) comes in: your business data is maintained centrally once and delivered to humans and machines in a clean, structured form — from a Single Source of Truth, without visual marketing and machine-readable facts drifting apart. Honest stays honest: no technology can force an AI to always cite correctly — the models are too probabilistic for that. But if you present your facts unambiguously and with evidence, you drastically lower the risk of misrepresentation and increase the chance that "AI knows you" becomes "AI knows you correctly". ## Sources - **AI Search Has a Citation Problem** — Tow Center for Digital Journalism, Columbia University, March 2025. [cjr.org](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php) - **AI search engines fail to produce accurate citations in over 60% of tests** — Nieman Lab, 2025. [niemanlab.org](https://www.niemanlab.org/2025/03/ai-search-engines-fail-to-produce-accurate-citations-in-over-60-of-tests-according-to-new-tow-center-study/) - **GEO: Generative Engine Optimization** — Aggarwal et al., ACM SIGKDD 2024. [arxiv.org/abs/2311.09735](https://arxiv.org/abs/2311.09735) *Note: the study captures a single point in time — error rates for individual models change with new versions. The underlying challenge (uncertain sourcing, high confidence despite errors) holds regardless of the specific figure.* --- Source: https://btlabs.dev/en/posts/when-ai-cites-you-wrong Last-Modified: 2026-08-03T11:05:22.070Z Languages: [de](https://btlabs.dev/llms/de/posts/wenn-die-ki-dich-falsch-zitiert) · [it](https://btlabs.dev/llms/it/posts/quando-l-ia-ti-cita-male) See also: [llms.txt](https://btlabs.dev/llms.txt) · [ai.txt (Policy)](https://btlabs.dev/ai.txt) · [identity.json](https://btlabs.dev/identity.json)