The following three examples come from Rheinwunder projects and analyses in 2026, anonymised here.
Fact check for a large organisation (September 2026): we compared 450 AI answers (150 questions in ChatGPT, Claude and Gemini) with the content of the organisation’s website. Only 20% were fully correct, 37% correct but imprecise, 38% partly wrong and 2% wrong; the remaining 3% had no connection to the organisation. Around a third of the errors could be traced back to the organisation’s own website: information was only available in PDFs, missing altogether or contradictory across different pages. These are exactly the points that can be fixed.
Winery with an online shop (September 2026): when asked about the winery by name, the AI systems mentioned it in 21 of 24 answers. For general questions about wine recommendations from the region, only in 26 of 126, while a competitor appeared in 59. Gemini also gave a wrong address and wrong opening hours.
Log file analysis of a high-traffic website (August and September 2026): over ten days, we counted almost 29,000 requests from ChatGPT-User, an average of around 2,900 a day. This is an upper limit, because monitoring tools also trigger such requests. They included all 100 pages with the most Google clicks. 15% of Claude’s requests ended on 404 pages because the model had invented URLs. We created a redirect list for these.