Search for a small technical problem on Google, something like why a Python package refuses to install. There is a decent chance you will get an answer before you see the website that explained the fix. You read the instructions, copy the command, and close the tab. Job done. From the reader’s side, that sounds like a better search experience.
The person who wrote that explanation has a different problem. They spent time finding the error, testing the solution, cleaning up screenshots, and keeping the article updated when the software changed. Their page may have helped Google produce a useful answer, but the reader never needed to visit it. No visit means no ad impression, no newsletter signup, and no chance for that writer to become someone the reader follows.
This is now more than a complaint from unhappy publishers. On October 25, 2026, The Verge published a fresh look at how Google’s AI features have changed the relationship between Search and the open web. Earlier this year, researchers ran controlled experiments to see what happens when users get AI answers instead of normal search results. The numbers are uncomfortable. And the part I find more concerning is that Google and publishers can look at the same search market and reach completely different conclusions about whether things are going well.
Access without medium partner: Google’s Search Is Keeping the Answers
Google Changed the Deal Without Removing the Links
For years, the arrangement was easy to understand. A publisher wrote something useful, Google indexed it, and a person searching for that information could click through to the original website. Google made money from ads around search results. Publishers earned attention, and sometimes money, when the reader visited them. It wasn’t always a fair deal, and getting onto page one could turn into a miserable SEO contest. Still, there was a reason to keep writing useful pages.
Google began rolling out AI Overviews in 2024. Those summaries appear inside regular search results and give users an AI-written response with links to sources. In May 2025, Google brought out AI Mode, a more conversational search experience where users can continue asking questions without returning to a familiar list of websites. Those products aren’t identical. An Overview adds an answer to a normal search page; AI Mode can make the answer itself the main experience.
By August 31, 2026, Google’s own update for website owners said AI Overviews had more than 2.5 billion monthly active users and AI Mode had passed one billion. Those are company-reported audience figures, and they should not be added together because the same person may use both. Still, the reach is enormous. This isn’t a small experiment that publishers can simply ignore until next year.
The links haven’t disappeared. Google has made them more visible in several updates, adding previews and links inside AI responses. But the order of events has changed. The reader gets an answer first and may click later, if there is a reason to. That small change can be expensive for websites built around explaining ordinary questions.
And ordinary questions make up a lot of the internet. A recipe measurement, an error code, a camera setting, a deadline, a quick explanation of a finance term. People search because they want an answer, not because they have a special attachment to whichever website wrote it. I can’t blame someone for choosing the faster option. I do wonder who pays to keep producing those answers once the clicks start disappearing.
The Numbers Are No Longer Just Publisher Complaints
The first set of numbers comes from Pew Research Center. It studied browsing activity from 900 US adults in March 2025, covering 68,879 unique Google searches. Around 18% of the searches in its dataset produced an AI summary when the researchers later collected the results. So this was an early view of the feature, before some of Google’s 2026 changes.
Here is what Pew observed. When an AI summary appeared, people clicked a traditional search result in 8% of visits. Without a summary, they did so in 15% of visits. Links inside the AI summary itself received clicks in only 1% of those visits. People were also more likely to end their browsing session after a results page containing an AI answer: 26%, compared with 16% for pages without one.
Those percentages describe what the researchers observed. They don’t prove that an AI summary caused every missing click. Searches that bring up summaries can be different from searches that don’t. Someone asking for a quick definition behaves differently from someone trying to buy a graphics card or read an investigation. Pew’s sample was also limited to tracked US users. The research matters because it shows a real pattern, not because it gives one universal traffic-loss percentage for every site on the web.
By itself, the 8% versus 15% comparison would leave a big question unanswered: what if Google simply shows AI summaries on queries that never attracted many clicks anyway? Fair question. I thought that was the strongest defense of AI search for a long time. A person who searches for the time in Tokyo probably doesn’t need to open three websites to get a clock. But there is newer research that tries to separate the AI feature from the type of question someone asks.
Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University used a Chrome extension to randomly give participants different versions of Google Search. In their April 2026 working paper, revised in July, they reported that when an AI Overview would have appeared, showing it reduced outbound organic clicks by 39.8% compared with hiding it. The researchers also reported a 34.5% increase in zero-click searches under that condition.
That is a much more direct test. People were assigned different search experiences rather than only being observed after Google decided what to show them. The researchers reported no measurable improvement in users’ perceived search quality or ease of finding information. They also found no difference in engagement on the destination websites among people who did click. That last result matters because it challenges the easy assumption that every lost click was low-quality traffic nobody would miss.
None of this means a newspaper loses exactly 39.8% of its readers the minute AI Overviews appear. The experiment measured a defined behavior under particular conditions. Publisher traffic also changes because of news events, ranking updates, seasonality, social networks, and what people are interested in that week. If a study tells you the exact cause of a certain kind of lost click, it doesn’t automatically tell you the total loss across a whole publishing business.
A Second Experiment Made AI Mode Look Even Harder on Websites
In August 2026, researchers Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson, and Danaé Metaxa published another field experiment on arXiv. Their team recruited US Chrome users and compared normal Google Search with experiences that hid AI answers or routed searches into AI Mode. The behavioral sample included 1,100 people who searched during the experiment. They were observed over a short period, with seven days of assigned search behavior.
When users were pushed into AI Mode, the share of searches that resulted in a click to an outside website fell by 18.8 percentage points relative to normal Google Search. That is a change in percentage points, not an 18.8% relative decline. People assigned to AI Mode also conducted about 0.92 fewer Google search sessions a day. The researchers saw fewer users clicking news sites, Reddit, and Wikipedia as well. This was a strong result for a fairly short test.
The experience wasn’t universally loved, either. Participants assigned to AI Mode gave lower ratings for usefulness, satisfaction, control, and trust in information found through Google. Some liked getting quick answers. Others complained that finding the actual website they wanted became harder. It’s quite a funny outcome: a feature built to make search easier made some heavy Google users less happy with it.
There are limitations. The sample skewed younger and more educated than the whole US population, and seven days is not enough to know how people will behave after six months. Participants normally used AI Mode for only around 0.6% of their searches before the experiment, so forcing it to become their default tested a future-looking scenario rather than today’s voluntary behavior. And during the no-AI part of the study, Google changed its page structure, which meant the researchers’ extension initially hid only about half the AI Overviews it was supposed to hide. They adjusted their estimates, but that complication belongs in the story.
Still, two different research efforts point in the same direction. If an answer arrives before the link, fewer people need the link. The remaining question is whether Google can keep sending enough useful traffic, money, and attention to the websites doing the original work.
Google Says Search Still Sends Billions of Clicks
Google doesn’t accept the simple claim that AI Search is destroying the web. In an August 2025 post, Search head Liz Reid said overall organic click volume from Google to websites was relatively stable compared with the year before. She also said Google was sending slightly more “quality clicks,” meaning visits where users don’t return to Search quickly. Google’s argument is that some quick fact-check visits are disappearing, but people who want more information still go to websites.
I think that defense deserves to be taken seriously, even if it doesn’t settle the argument. Google sees a huge amount of search activity across countries, topics, and websites. A specialist news publisher only sees its own referrals. Total clicks could stay roughly flat if some types of websites gain traffic while others lose it. It is possible for Google’s overall numbers and a publisher’s declining traffic chart to be true at the same time.
The missing piece is the distribution. Google has not published enough comparable query-level data for outsiders to independently confirm every claim about improving click quality. A stable global click total doesn’t tell a 12-person publication why its guides now bring in fewer readers. It also doesn’t tell a reader whether Google’s answer was based on a page that can afford to remain online next year.
Google’s later actions show it knows publishers want more than reassurance. In 2026, the company introduced additional source previews, inline links, subscription labels, and more controls for website owners. It has also been testing ways for publishers to manage appearances in generative search features. Those are real product changes, although better-looking links don’t automatically restore the traffic that old search results once delivered.
One more awkward point: Google can now answer some questions in a space where it also sells advertisements. An independent site generally needs people to visit before it can earn ad money. The financial incentives are different. That isn’t proof that Google deliberately wants publishers to fail, but it helps explain why the companies disagree so sharply about what a healthy search result looks like.
Small Sites Have a Different Problem From Large Newsrooms
The UK Association of Online Publishers published research in 2026 using six months of traffic data from eight major publishing groups. The sample covered 10.8 billion pageviews across participating businesses. Between the first and second quarters of 2025, Google organic search referrals fell by 7.1% across the group. News publishers saw a smaller average quarterly fall, roughly 5%, while consumer and business-to-business titles saw a fall closer to 18%.
The sample is big, but it is still a selected group of established UK publishers. It cannot stand in for every independent blog, local news site, or Medium writer. The study itself makes that limitation clear. Sites that depend on high-volume search traffic and programmatic ads may have less room to absorb losses than publications with loyal subscribers.
A different data set from Define Media Group, reported in March 2026, looked at 64 sites and found a 42% fall in organic Google Search clicks over its comparison period. Meanwhile, breaking-news traffic across Search, Discover, and Google News increased 103% in that portfolio from November 2024 into early 2026. Those are measurements from the company’s own clients, not a general census of the web. But the split is interesting: content that answers a question anyone can summarize may be under more pressure than reporting that tells people something they couldn’t have known yesterday.
For a small publisher, this changes the economics of an ordinary tutorial. An article may take an afternoon to research and another afternoon to write. Maybe it earns enough from ads or affiliate links to justify that time when Google sends steady visitors. Cut those visits and the calculation changes. The author stops updating the page, or decides the next guide isn’t worth producing. The first loss shows up as fewer clicks. The second loss is harder to see because the next useful article never gets written.
A large publisher can try paid newsletters, subscriptions, licensing agreements, live events, or direct sponsorships. Someone running a two-person tech blog often has none of that ready. And moving everything to video or social media isn’t a magic escape route. Those platforms change their algorithms too, and sometimes faster than Google does.
Google Has Started Paying Some Publishers. That Creates Another Problem.
On October 15, 2026, Google said more than 400 publications were participating in its News AI pilot. The program includes payment for delivering and using enhanced content in new AI news features. Google named partners including CNN, Axios, the Financial Times, Forbes, and smaller regional publishers. It describes the program as a way to create better news experiences with attribution and links back to publications.
This is a real change from expecting every publisher to survive entirely on whatever traffic an AI answer sends. Payment for content can help publishers finance reporting even if readers get a short summary before opening the full story. The catch is that a pilot involving 400 publications doesn’t create a business model for the rest of the internet. Millions of websites sit outside those partnerships, and there is no published universal payment rate that lets a small writer estimate what their work is worth to AI Search.
And a licensing deal can create strange incentives. Large publications can negotiate. A niche writer with a brilliant database troubleshooting guide probably can’t phone Google’s partnership team and agree on a price. If the biggest brands get paid and the long tail gets cited occasionally, the open web may become less diverse even while Google’s search pages look busier.
There are legal fights around this too. The Verge’s October 25 reporting notes publisher lawsuits and regulatory attention, along with cases where judges rejected certain claims. That is why I wouldn’t announce that Google has been found legally responsible for all lost traffic. The commercial and legal arguments are still unfolding, and the outcome will probably differ across countries.
A Better Strategy for Writers Is More Specific Than “Do SEO”
If you publish on Medium or maintain your own website, I would change what you choose to write before worrying about a new keyword trick. General explainers are easy for an AI answer to compress. A post titled “What Is CUDA?” can still help beginners, but Google has many ways to answer it in a few sentences. A first-hand comparison of local AI performance on an RTX 5090, with model versions, quantization settings, power readings, and the results that failed, gives readers a stronger reason to open the article itself.
That doesn’t mean every article needs a lab or expensive hardware. Original value can be an interview with someone who ran into the problem, a carefully checked timeline of a confusing company announcement, a repair diary with photographs, or a clear explanation of why two popular claims contradict each other. Readers visit when there is something they can’t get from a generic answer box. And if you can show the work, even better.
For an existing site, open Google Search Console and look for pages where impressions remain steady but click-through rates have fallen. Compare similar date ranges and separate branded queries from questions or how-to searches. Then look at whether AI Overviews actually appear on those queries. The change might come from search features, or it might come from rankings, competing pages, a seasonal dip, or a new search habit. Don’t call every traffic decline an AI problem just because that is the fashionable explanation.
Build a way for interested readers to come back without asking Google for permission each time. An email list is useful because the subscriber has asked to hear from you directly. A Medium follower relationship, a small discussion community, an RSS feed, or a simple series of articles can also create repeat visits. The goal is to make the next reader less dependent on discovering your work through a one-off search query.
And keep the boring measurements. Track which articles bring readers back, not only which ones got a search spike. Record direct visits, newsletter signups, returning readers, and revenue per article if you earn money from publishing. A post that attracts 2,000 readers who remember your work may be more valuable than one that briefly attracts 20,000 visitors looking for a single copy-paste answer. Those are example numbers, not a universal formula, but the difference in reader intent matters.
There is one easy trap here. Publishers might respond by producing twice as many AI-written pages to replace the traffic they lost. I think that is the wrong bet for most independent writers. If Search is learning to answer routine questions itself, flooding it with more routine answers gives Google more material without giving readers a stronger reason to know who wrote it. You end up working harder for the exact kind of click that is disappearing.
The Next Search Problem Belongs to Google Too
Google has a decent short-term product argument: someone asks a question, gets an answer, and moves on. As a user, I like that when the answer is correct. I don’t want to open six pages full of popups just to find a command that takes ten seconds to run. Publishers helped create some of that bad experience themselves, and pretending otherwise doesn’t help their case.
But the long-term bargain is more difficult. AI Search needs original material to keep answering current questions. Someone still has to investigate a broken release, test a new chip, attend a court hearing, or find the bug that wasn’t in last year’s documentation. If those people cannot earn money or an audience from publishing the result, the supply of fresh useful information becomes harder to maintain. That risk is a reason to watch publisher economics, not proof that the web will disappear.
The latest experiments tell us that fewer users click when AI answers take over more of Search. Google’s own numbers say there is still a huge amount of web traffic and a growing audience for AI features. Both observations deserve a place in the discussion. The hard part is finding a model where getting an instant answer doesn’t quietly remove the reason someone wrote the answer in the first place.
Next time you search for a tiny Python fix and Google gives you the command immediately, it may be worth opening the source link anyway. There could be a note below the first solution explaining why the command fails on Python 3.13. That little detail is often the difference between an answer that looks useful and one somebody took the trouble to test.