Do people actually trust AI answers?
Less than you would think, and more than they should. Both halves of that are measured, and the space between them is where your customers now live.
Start with the half that surprises people. The largest direct experiment on this question ran about twelve thousand queries across seven countries, then put a randomized experiment in front of a sample built to represent the U.S. population. On average, people trusted AI-generated search results less than traditional search results. Not more. That work is a preprint, not yet peer reviewed. It is also the largest direct test anyone has run on the question. The picture of consumers swallowing whatever the machine says is not what the data shows. People arrive at an AI answer with their guard partly up, the same way they learned to squint at ads and sponsored results over twenty years of using the web.
Now the half that should worry everyone. The same experiment found that reference links and citations significantly increased trust in AI answers, even when the citations were wrong or entirely made up. Read that again slowly. The trappings of evidence did the work of evidence. A fabricated footnote raised trust about as well as a real one, and in the same study, the readers who trusted most verified least. And the study found one more thing, the strangest of the three: when the AI displayed its own uncertainty, saying in effect how confident it was, people trusted and shared the answer less, whether the stated confidence was high or low. Honesty about limits was punished. The confident, footnoted answer wins the reader, and its footnotes go unexamined.
I spent years in my doctoral work on a related question, which is why this result reads as familiar rather than shocking to me. Psychology has known for decades that ease does a quiet kind of persuading. When information arrives fluently, cleanly written, well organized, effortless to take in, people rate it as more true and more trustworthy than the same information delivered with friction. The feeling of easy reading gets mistaken for the quality of the thing being read. An AI answer is the most fluent format ever built for delivering information. One tidy paragraph, no ads, no navigation, no choosing between ten sources. The old results page made you do the work of judging, and the work itself kept some skepticism alive. The answer format removes the work. What I will say carefully, because no study has tested this exact chain for AI answers yet, is that this is my read on why the citation effect is so strong: the whole package feels finished, so checking it feels unnecessary.
So what does the honest picture look like? People have not stopped being skeptical. They have moved their skepticism to the wrong checkpoint. They discount the AI channel in general, then accept the individual answer in front of them because it looks and feels like a finished, sourced piece of work. Verification did not disappear. It got redistributed, away from the answer itself and onto surface signals like citations, formatting, and confidence, which are exactly the signals that can be manufactured.
If you run a business, this cuts two ways, and both matter.
The first is defensive. When an AI answer says something wrong about your business, your prices, your hours, whether you are even open, the people reading it are not fact-checking it. The citation effect means a wrong answer with sources attached gets believed. There is no truth police coming, and whatever sits in the record tends to get repeated. The only real countermeasure is upstream: keep the checkable record of your business accurate and consistent everywhere a machine might read it, so the answer assembled about you is built from facts you actually control. I have written before about what that record needs to contain, and this is the reason it matters more than any promotional work you could do.
The second is a line I will not help anyone cross. If surface signals of evidence manufacture trust, then stuffing your pages with citation-shaped decoration is an obvious temptation, and parts of this industry are already selling it. Do not. The systems policing this are the same ones that spent two decades learning to catch manufactured signals. Engineered, answer-shaped content is already treated as spam by at least one major engine. The entire lesson of the last twenty years of search is that fake signals work right up until an update makes them a liability. The durable version is slower and duller: be the business about which the accurate, checkable, well-sourced answer can be written, because real evidence and manufactured evidence look identical for a while and then very suddenly do not.
Do people trust AI answers? They distrust the channel, trust the format, and rarely check. That gap is not your opportunity to exploit. It is your responsibility to survive, and the businesses that treat it that way will still be standing when the checking gets better, which it will.
This essay is part of The Discovery Layer: How Businesses Get Found in the Age of AI, a book I'm writing in public. Each essay becomes a chapter, and the book expands them with the full evidence. Get on the list and I'll tell you when it ships.