AI answers are making people faster at reaching decisions, and measurably worse at making them. Two independent studies published in 2025, using different methods and different populations, reached the same conclusion: people who get answers from AI summaries learn less, produce lower-quality advice, and click on almost nothing they are shown. For anyone trying to be found and trusted in AI-powered search, that convergence is worth examining closely.
What Is “Path Metadata” and Why Do AI Answers Lose It?
Before AI answers existed, the journey toward information was itself informative. Three conflicting sources meant a topic was contested. A search that returned nothing meant the territory had not been mapped yet. These signals, what Search Engine Journal contributor Duane Forrester calls “path metadata” (the byproduct of traveling through information rather than landing directly at a conclusion), shaped how firmly you committed to what you found. AI answers strip those signals out entirely, returning the same confident prose whether the evidence behind them is deep or nearly absent.
The loss is not about answer quality alone. It is about calibration, meaning knowing how much to trust what you now know. You arrive at the conclusion without the means to evaluate it. That distinction matters for publishers, because it changes who is arriving at their content, and in what state of mind.
What Do Studies Say About AI Answers and What Users Actually Learn?
Two research streams, published within months of each other, point the same direction. Together they provide the clearest empirical picture yet of what AI-assisted research does to knowledge quality.
The first is a controlled experiment. Shiri Melumad and Jin Ho Yun, marketing professors at Wharton, ran seven experiments with 10,462 participants and published the results in PNAS Nexus in October 2025. Participants learned about everyday topics, planting a vegetable garden or spotting financial scams, from either an AI summary or standard Google links. Those who used AI came away knowing less, even when both groups saw identical underlying facts. They spent less time engaging with the material. The advice they wrote afterward was sparser, less original, and less likely to be acted on by the people who read it. A version where the model provided live web links alongside its answer changed nothing: participants did not click them once the summary arrived.
The second is a passive behavioral study. Pew Research Center tracked the real browsing of 900 U.S. adults across 68,879 Google searches in March 2025. When an AI Overview (Google’s AI-generated summary block at the top of results) appeared, users clicked a normal search result on 8% of visits, compared to 15% when no summary appeared. They clicked a source cited inside the AI summary on roughly 1% of visits. They ended the browsing session entirely on 26% of pages with an AI summary, versus 16% of pages without one. Pew notes this is association rather than proven cause, and covers one month, U.S. users, Google only.
A 2025 paper from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real AI uses at work and found that greater confidence in the AI predicted less critical thinking. The study is self-reported and correlational, and one co-authoring institution sells a generative AI product, context worth noting. A 2026 exploratory study by Dirk Lewandowski on information regret points in the same direction from a smaller sample.
Why the Old Correction Mechanism No Longer Works
Under the previous model of search, a wrong answer was survivable. Someone read a bad summary, kept going, landed on the source page, and replaced the wrong version with the right one. That repair was free, automatic, and ran on ordinary user curiosity. At a 1% source-click rate, it effectively stops.
As Forrester describes it, a correction mechanism that cost nothing and fired in seconds has been replaced by one that costs real effort and fires slowly, if it fires at all. A misrepresentation of your brand or expertise used to be temporary. People would click through, find the real version, and update their understanding. Now it persists, because almost no one clicks to check.
The asymmetry is the expensive part. The old repair asked nothing of the publisher. Whatever replaces it requires actively publishing evidence, waiting for it to be crawled, retrieved (pulled from a data store to answer a specific query), and weighted, with no confirmation that any of it worked.
What Does This Mean for Content Strategy?
A decade of content strategy assumed a staircase: introductory explainers at the top for newcomers, comparisons in the middle, depth at the bottom for the well-read visitor. The Wharton and Pew findings together suggest the top of that staircase now happens before anyone reaches a publisher’s site. People arrive already past the basics, but carrying a confidence level that outpaces their actual understanding.
Forrester identifies the specific mismatch this creates. Beginner content talks down to someone who already believes they know the fundamentals. Advanced content assumes vocabulary they can repeat but have not earned. A site misses in both directions at once, using assets built over years.
His prescription: content only a publisher can produce, original data, firsthand experience, specific expertise, was positioned as the depth layer. Given that the basics now often happen before anyone arrives, that material may need to serve as the front door as well. The cleanest, most structured, most plainly explanatory pages are the easiest for a model to absorb and restate, which means they are often the ones a newcomer no longer needs to visit. That work still happens. It happens inside the AI answer now.
What This Means for AI-Search Visibility
Here is the implication that most traffic-focused analysis skips: if nearly no one clicks citations inside an AI answer, the value of being cited is not referral traffic. It is trust-by-proxy. An AI cites your brand, a reader does not click, but they act on the answer carrying your credibility without ever visiting your page. That is a fundamentally different relationship between content and audience than any previous channel created, and standard analytics will not capture it.
That shift cuts two ways. A brand cited consistently in AI answers for a category question builds association that shapes decisions even when no visit occurs. A brand absent from those answers, or misrepresented in them, has lost the correction mechanism it once received for free. No one is clicking through to find the real version.
The Melumad and Yun finding about advice quality adds a dimension that most AI-search discussions miss entirely. It is not only about whether your brand appears in an AI answer. It is about the quality of understanding that travels with the citation. An AI summary that flattens expertise into a generic paragraph produces readers who are, as Forrester puts it, “confidently underinformed.” That reader will be harder to persuade than an uninformed one, because they believe the research is complete.
In Hingewise’s assessment, this is a share-of-model problem (the measure of how often AI mentions your brand versus competitors) with a quality dimension attached. Appearing is necessary. Appearing with enough specificity that the AI summary carries real signal, rather than language that could describe any company in your category, matters equally. The research suggests the pages built to bring strangers up to speed are the ones models absorb most readily. Those pages now function as AI training and retrieval material whether or not a visitor ever lands on them directly.
The clearest practical signal: check which of your pages an AI currently uses to describe you. If those pages are accurate and specific, you are in reasonable shape. If they are generic, or if the AI describes a competitor instead, the correction mechanism that used to fix that for free is no longer running.
Quick Checklist: Before You Audit Your Content for AI Visibility
- Test what ChatGPT, Gemini, and Perplexity currently say about your brand and category. This is your de facto front door, not your homepage.
- Check whether that AI summary is accurate, specific, and differentiated, or generic enough to describe any competitor.
- Identify your cleanest, most structured pages. These are the ones models absorb most readily; verify they contain your actual differentiators.
- Confirm that original data, firsthand experience, or specific expertise appears on pages a crawler can easily reach and parse.
- Note where your brand is absent from AI answers in your category. Absence is not neutral if a competitor consistently fills that space.
- Consider how content written for a “curious researcher” reads to someone who already has an AI summary and believes they are done researching.
- Track branded search volume and direct navigation alongside AI citation monitoring. If AI influence is growing, those numbers should move even when click-through from AI does not.
What to watch going forward: whether platform changes to AI search interfaces alter the roughly 1% citation-click rate meaningfully. If that number rises, the referral value of AI citation increases. If it holds near zero, the case for treating the AI answer itself as the primary audience, rather than a funnel stage before the primary audience, only strengthens.
Sources: Duane Forrester, Search Engine Journal (2026); Melumad and Yun, PNAS Nexus (October 2025); Pew Research Center (March 2025); Microsoft Research and Carnegie Mellon University (2025); Lewandowski, information regret study (2026).
Sources
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.