AI search trust problem widens
A recent study found Google’s AI-driven Overviews get about 10% of strictly verifiable questions wrong, a mistake rate that becomes meaningful when scaled to billions of queries per hour. That credibility gap is already prompting competitors and vendors to build tools for brands to monitor generative search visibility as trust and attribution become business problems. (mobilesyrup.com, itbrief.asia)
Google built AI Overviews to answer your question before you click a link, like getting the summary of a book on the cover instead of opening the first page. By May 2025, Google said AI Overviews had reached 1.5 billion monthly users in 200 countries and territories. (blog.google) That scale turns a small error rate into a very large output problem. A recent analysis reported by Ars Technica, based on testing described as using more than 4,000 strictly verifiable questions from the SimpleQA benchmark, found Google’s AI Overviews were wrong about 10 percent of the time. (arstechnica.com) Google’s own pitch for AI Overviews is speed and convenience, not just novelty. In its May 20, 2025 product update, Google said the feature appears when its systems judge it helpful and claimed it drove more than a 10 percent increase in usage for the kinds of queries that show AI Overviews in the United States and India. (blog.google) That changes what a search result is. Semrush described AI Overviews as a shift from Google acting like a list of links to Google acting like an answer engine, with summaries appearing above organic results and often reducing the need to click through to websites. (semrush.com) Once the answer is generated on the page, the old web bargain starts to wobble. Publishers used to trade content for traffic, but an artificial intelligence summary can absorb the content, answer the question, and send fewer visitors to the original source. (semrush.com) That is why the next business fight is not only about ranking on Google’s first page. It is about whether a brand, publisher, or review site gets cited inside the machine-written answer that the user sees first. (blog.google) Trustpilot moved directly at that problem on April 7, 2026. Its new AI search analytics product says it shows, in real time, how often a company’s Trustpilot profile is cited by ChatGPT, Claude, and Perplexity, with data backfilled to November 2025. (trustpilot.com) Trustpilot is selling that visibility as a trust product, not just a marketing dashboard. In its launch announcement, the company said 58 percent of consumers now turn to generative artificial intelligence for recommendations and said brands not recognized as “the answer” by AI engines could see organic traffic drop by 20 percent to 50 percent. (prnewswire.com) The company is also arguing that review data has become raw material for machine answers. Trustpilot said it hosts more than 361 million active reviews, removed 7.8 million fake reviews in 2025, and detected 91 percent of those fake reviews automatically. (prnewswire.com) So the trust problem now has two layers. Users need the answer itself to be correct, and businesses need to know whether the answer was built from reliable sources or from whatever the model found easiest to summarize. (arstechnica.com, prnewswire.com) Google can keep improving the model, and vendors can keep building citation trackers, but neither fixes the core tension. The faster search becomes an instant-answer machine for 1.5 billion monthly users, the more every wrong sentence and every missing source turns from a product flaw into an economic one. (blog.google, arstechnica.com)
