An LLM takes present-you, the version with a question, and hands back future-you, the version with a decision.
That is the trick, and it is a real one. Go back far enough, and answering a serious question meant a library. You looked up books, read them, took notes, found references inside those books that sent you to other books, and after some number of days you knew enough to decide. Search compressed that into hours. You asked, you got results, you followed the clues inside those answers into better questions, and you built a picture until the picture was vivid enough to act on. An answer engine compresses it again, into seconds. Same starting point. Same destination. Almost no travel.
Every limit that applied to the older versions still applies to this one. The library either had the right books, or it did not, and you had no way to check. Google either surfaced the right sources or it did not, and it introduced a problem libraries mostly avoided, because getting a book published took effort and oversight while publishing a web page took neither. Errors, inventions, and confident nonsense scaled past anything a person could sort. You decided anyway. Everyone did.
So the speed is new. The uncertainty underneath it is not. What is new is quieter, and it is the part that’s the concern for anyone who publishes for a living.
The journey was never only moving you toward an answer. It was telling you, the entire time, how much that answer was worth. Three books and no journal articles meant the topic was thin. Sources contradicting each other meant it was contested. A search that returned nothing meant you were asking about territory nobody had mapped yet. A question that took four hours felt different at the end than one that took four minutes, and that difference was information. Nobody sat down and consciously read those signals. They were simply present, in the record, shaping how firmly you committed to what you found.
Call it path metadata. It is a byproduct of travel rather than a feature anyone built; it never survives into the destination, and until recently it did not need to, because you could not reach the destination without generating it along the way.
An answer engine returns a conclusion in the same confident prose whether the evidence behind it was deep or nearly absent. The compression is lossy in one specific direction: it strips out exactly the signals you would have used to judge what you were handed. You arrive at the answer without the means to evaluate it.
The Research Caught Up This Year
For a long time this was arguable but not demonstrable, but that changed.
Two marketing professors at Wharton, Shiri Melumad and Jin Ho Yun, ran seven experiments with 10,462 people and published the results in PNAS Nexus in October 2025. Participants learned about ordinary topics (planting a vegetable garden, spotting financial scams) either from an AI summary or from standard Google links. Then they wrote advice for someone else based on what they had learned.
The people who used the AI came away knowing less. That was true even when the facts placed in front of both groups were identical. They spent less time engaging with what they got. And the advice they wrote afterward was sparser, less original, and less likely to be taken by the people who read it.
Then the finding that closes the obvious escape hatch. The researchers ran a version where the model supplied live web links alongside its answer. Participants did not go look. Once the summary arrived, the sources sitting right beside it stopped being interesting.
Pew Research Center found the same behavior in the wild. They tracked the real browsing of 900 U.S. adults across 68,879 Google searches in March 2025. When an AI summary appeared, people clicked a normal search result on 8% of visits, against 15% when no summary appeared. They clicked a source cited inside the summary on roughly 1% of visits. They ended the browsing session entirely on 26% of pages with a summary, against 16% of pages without one. Pew is careful to call this association rather than proven cause, and it covers one month, US users, Google only.
A controlled experiment and a passive tracking study, different methods, different people, same destination. That convergence is why I am willing to build on it.
The last piece predates the technology. In 2015, three Yale researchers ran nine experiments and found that searching the internet inflated how much people believed they knew. They confused access to information with their own understanding of it. The effect appeared even when both groups saw identical content. It also appeared after searches that turned up nothing at all.
That last detail complicates my own argument, so let me deal with it rather than steer around it. If people were already confusing access with understanding a decade ago, and already grew more confident after finding nothing, then I cannot claim the old journey taught good judgment, as it did not. What the journey left behind was friction, elapsed time and visible variety in the sources, present in the record whether or not anyone used them deliberately. Melumad and Yun show those things were doing work. Remove them and the deficit turns up in what people produce.
One more, with a disclosure attached. A 2025 paper from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real uses of AI at work. Greater confidence in the AI predicted less critical thinking. Greater confidence in their own ability predicted more of it. The study is self-reported and correlational, and I will note that one of the co-authoring institutions sells a generative AI product. Dirk Lewandowski’s 2026 study of information regret points the same direction from a smaller, more exploratory sample.
The Correction That Used To Happen For Free
Here is where this stops being a story about users and becomes a story about your work.
Under the old model, a thin or wrong answer was survivable, because the journey repaired it. Someone read a bad summary, kept going, landed on your actual page, and swapped the wrong version for yours. Nobody designed that. It cost nothing. It happened millions of times a day, and it functioned as the immune system of the entire information economy.
At a 1% source-click rate, it does not fire. The time machine not only skips the journey, it skips the repairs that used to happen along the way.
I wrote recently about what happens when a model has nothing on your company and describes a competitor instead, and about why no content audit will ever surface it. That piece was about why the failure is invisible. This one is about why it now lasts. The misrepresentation used to be temporary. Now it stays put.
That asymmetry is the expensive part. The old repair was free, automatic, and ran entirely on somebody else’s curiosity. Whatever replaces it has to be paid for. You publish the evidence, then wait for it to be crawled, retrieved, weighted or trained on, with no control over the timeline and no confirmation that it worked. A correction mechanism that cost nothing and fired in seconds has been replaced by one that costs real money and fires slowly, if it fires at all.
Citation Is Presence, Not Traffic
If roughly one visit in a hundred produces a click, then treating AI citation as a referral channel values it wrong. The value was never the visit. The value is being inside the answer that a person acts on, which means the thing you measure has to be the answer itself rather than the trickle that escapes it.
Your Inbound Lead Is Not Uninformed
The person who lands on your site or your sales call is not uninformed. They are confidently underinformed, and content written to educate a curious researcher does not work on someone who believes the research is already done.
That claim is mine rather than the studies’. None of those papers examined marketing content or funnel structure. But follow what they found into your own asset library, and the consequence is difficult to dodge.
A decade of content strategy assumed a staircase. Definitional explainers at the top for the person starting out. Comparisons in the middle. Depth at the bottom for the person who had done the reading. I have argued before that the top of that staircase now happens before anyone reaches you, and that the people who do arrive are further along the decision journey than they used to be. I still think that is right. What I had wrong was treating further along and better informed as the same thing. They are not. The person arrives carrying the confidence of someone who finished the research alongside the actual depth of someone who read one paragraph, and those two things pull in opposite directions.
Your beginner content now talks down to them, and being talked down to ends a visit quickly. Your advanced content assumes a vocabulary they can repeat but have not earned, and that ends the visit too, more politely. You miss in both directions at once, using assets you spent ten years and a large budget building.
Deleting the 101 material is not the answer. It still feeds what the models say, whether or not you get credit for it. What has to change is placement. The material only you can produce was positioned as the depth layer, the thing a reader reached once the basics were out of the way. The basics now happen before anyone arrives, which means the content a model cannot perform has to do the job of the front door as well. I made the case a while back for where content defensibility comes from, and none of that changes. Where it sits does.
The Loop Nobody Funded On Purpose
Notice which of your pages a model performs best with. The cleanest, most structured, most plainly explanatory ones, the pages built to bring a stranger up to speed, are the easiest for a model to absorb and restate.
Which means the pages you built to greet a newcomer are the pages a newcomer no longer needs to reach. That work still happens. It happens somewhere else now, and the person arrives on the far side of it, already past the part you wrote for them.
I am not arguing you should stop producing it. Being missing from what the models read is worse than being absorbed by it. But you should stop expecting a decade-old content stack to greet an inbound lead who has already been fast-forwarded past its opening.
The Mirror
One last thing, and it points back at us.
You are also a user of these systems. The competitive analysis, the strategy deck, the recommendation you are about to hand a client or a board, if any of it came out of a model’s synthesis, then it is the sparser, less original output Melumad and Yun measured, and you are more certain about it than the process earned. The advice their participants wrote was less likely to be adopted by the people who received it. Those recipients could not say why. They just found it thin.
The time machine works. It genuinely takes you from question to decision in seconds, and most of the time the decision is fine. But it drops you somewhere without telling you how far you traveled or what you went past. Understanding how the system operates underneath is the only thing that restores any of that. Not because the answers are wrong. Because you cannot tell.
More Resources:
- When AI Has Nothing On Your Company, It Describes Someone Else
- How To Measure AI Search Visibility
- Study Confirms Google AI Overviews Cut Organic Clicks 38%
This post was originally published on Duane Forrester Decodes.
Featured Image: showcake/Shutterstock
