AI Answer Engines Strip "Path Metadata," Leaving Users Confidently Underinformed, Research Shows

A cluster of recent studies suggests that large language models (LLMs), by compressing the research journey into seconds, systematically remove the signals users once relied on to judge the quality of an answer — a phenomenon Duane Forrester, founder and CEO of UnboundAnswers.com, calls the loss of "path metadata."

The Research

The most controlled evidence comes from Wharton marketing professors Shiri Melumad and Jin Ho Yun, who published findings in PNAS Nexus in October 2025. Their study ran seven experiments with 10,462 participants who learned about everyday topics either from an AI summary or from standard Google links, then wrote advice for a third party based on what they had learned. According to the article, participants who used AI "came away knowing less," spent less time engaging with the material, and produced advice that was "sparser, less original, and less likely to be taken by the people who read it." The deficit persisted even when both groups were shown identical underlying facts. Critically, when the model supplied live web links alongside its answer, participants still did not follow them: "Once the summary arrived, the sources sitting right beside it stopped being interesting."

A separate passive-tracking study by Pew Research Center followed the real browsing behaviour of 900 U.S. adults across 68,879 Google searches in March 2025. When an AI summary appeared, users clicked a standard search result on 8% of visits, compared with 15% when no summary appeared. Clicks on sources cited inside the summary occurred on roughly 1% of visits. Users ended their browsing session entirely on 26% of pages that included a summary, versus 16% of pages without one. Forrester notes that Pew "is careful to call this association rather than proven cause, and it covers one month, US 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. Forrester flags that the study "is self-reported and correlational" and that one co-authoring institution "sells a generative AI product." A 2026 study by Dirk Lewandowski on "information regret" pointed in the same direction from a "smaller, more exploratory sample." An earlier 2015 study by three Yale researchers found that searching the internet alone "inflated how much people believed they knew," with the effect appearing even after searches that returned nothing.

What Publishers Lose

Forrester argues that the old model of search contained a self-correcting mechanism: a user who encountered a thin or wrong answer would keep browsing, eventually land on a publisher's page, and replace the wrong version with the correct one. At a roughly 1% source-click rate, he writes, "it does not fire." He describes the original correction loop as "free, automatic, and ran entirely on somebody else's curiosity," and argues it has been replaced by a mechanism that "costs real money and fires slowly, if it fires at all."

On the question of AI citation value, Forrester contends that treating AI citation as a referral channel "values it wrong," because "the value is being inside the answer that a person acts on" rather than the volume of clicks that follow.

Implications for B2B Content Strategy

Forrester extends the findings to B2B marketing funnels. He argues that inbound visitors now arrive "confidently underinformed" — carrying the confidence of someone who has finished researching, but the actual depth of someone who "read one paragraph." He notes that none of the academic papers examined marketing content or funnel structure directly, and characterises his funnel conclusions as his own interpretation rather than findings of the studies.

His practical recommendation is a repositioning of content: material that only a brand can produce — proprietary data, original analysis, first-hand experience — needs to function as the entry point rather than a depth layer, because introductory content is the easiest for models to absorb and restate in place of a direct visit. He also turns the argument on practitioners themselves, noting that any strategy output generated through an AI synthesis carries the same deficit the studies document: "you are more certain about it than the process earned."