Google’s answer engine faces a trust problem

Google Search used to make a simple promise. It would point you somewhere. Now it increasingly answers you itself, and that changes the standard it has to meet. A ranking system can be messy as long as the good links are in the pile. An answer engine has to be right, or at least right often enough that people cannot tell when it slips. A new analysis suggests Google is not there yet. Using a standard factual benchmark, researchers working with The New York Times found that AI Overviews answered 85% of questions correctly in October and 91% in February after a model upgrade, a real improvement that still leaves a very large error surface inside a product Google serves at planetary scale (searchengineland.com). That scale is the whole story. Google says it now handles more than 5 trillion searches a year, and it has expanded AI Overviews to 1.5 billion monthly users across 200 countries and territories. In Google’s biggest markets, including the United States and India, the company says the feature has already increased usage for the kinds of queries that trigger it by more than 10% (blog.google, blog.google). Even if a 90%-plus system sounds good in a lab, multiplying that miss rate across a search engine this large turns occasional mistakes into a structural quality problem. The trouble is not just wrong answers. It is answers that look sourced when the sourcing does not really hold up. The same analysis reported that more than half of the correct February responses were “ungrounded,” meaning the cited links did not fully support the claim being made. That figure had worsened from 37% in October to 56% in February. So the product got more accurate on the benchmark while becoming harder to verify from the evidence it showed users. That is a nasty trade. It teaches people to trust the summary more than the sources under it (searchengineland.com). Google has been arguing for nearly two years that AI Overviews are not just a chatbot bolted onto search. The company says the feature is tied into its ranking systems, built to surface information backed by top web results, and meant to avoid the free-floating hallucinations people associate with standalone AI assistants. When odd answers went viral after the U.S. rollout in May 2024, Google said many screenshots were fake and that the real failures often came from query misinterpretation, web ambiguity, or thin source material rather than pure invention (blog.google). That defense matters less than it used to. Once Google puts a synthesized answer at the top of the page, users do not experience a taxonomy of error. They experience confidence. And Google is not merely trying to make those answers useful. It is trying to make money from them. The company now places ads above, below, and in some cases within AI Overviews. Google says ads inside AI Overviews are available in English on mobile and desktop in the U.S. and several other countries, and that both the query and the AI-generated summary help determine which ads appear. The company frames this as a way to meet users in “new moments of exploration.” It is also a sign that the answer box is no longer just a search feature. It is inventory (support.google.com, blog.google). That commercial layer is what makes the timing of Turkey’s move feel less like a side story and more like the same story in another language. On April 3, Turkey’s Competition Board said it had opened an investigation into Alphabet and related Google companies over billing practices and other commercial conduct tied to Google’s online advertising services for advertisers and agencies. The question is whether those practices violate Turkish law (finance.yahoo.com). Search is becoming a machine that summarizes the web, decides what evidence counts, and sells placement around the answer while regulators ask how the billing works.