What AI Overviews Did to Organic Traffic

When Google started answering questions at the top of the results page, the worry was simple. If the answer is already sitting there, why would anyone click. That fear arrived before the evidence did, and for a while the debate ran on anecdote and screenshots. Enough real data has now come in to settle the broad shape of it. The worry was right for some kinds of pages and overblown for others, and the gap between those two groups turns out to be the whole story. What follows is what the studies actually found, why the loss landed where it did, what AI answers reach for when they cite a source, and what a team should do about it.

The shift in how people search

The behavior changed before the traffic did. People stopped treating search as a list of doors to walk through and started treating it as a place that talks back. A query that once returned ten blue links now returns a written answer first, with the links pushed below it or folded behind a citation. For a large class of questions that is enough, and the searcher never scrolls. The act of searching used to end on a website. Now it often ends on the results page itself, and the website is something the reader consults only when the summary leaves a doubt.

This is not only a Google story. The same habit is forming inside ChatGPT, Perplexity, and every assistant that reads the web and summarizes it. Across all of them the interaction has the same shape. A person asks in plain language, a model reads a handful of pages on their behalf, and the reader gets a synthesis instead of a set of options to sort through. The result is a layer between the searcher and the website, and that layer now does work the website used to do. It reads, it filters, and it decides which few sources are worth surfacing. The open question was never whether this would cost clicks. It was how many, and to whom.

The clicks that disappeared

The clearest read comes from Pew Research Center, which followed the real browsing of thousands of people across roughly 68,000 searches. On searches that showed an AI overview, people clicked through to a website on about 8% of visits. On searches without one, they clicked on about 15%. That is close to a 47% drop in the chance of a click once the answer sits at the top. The strength of the Pew work is that it watched what people did rather than what they said, and it did so at a scale large enough to trust.

A separate randomized field study, where some people were shown overviews and others were not, came at the question from the cleaner experimental angle and landed in the same neighborhood with a click loss around 38%. The gap between 38% and 47% is worth sitting with rather than averaging away. Different methods, different query mixes, and different time windows will move the number, and a young field measured by a handful of teams should be expected to disagree at the edges. What does not move is the direction. Every serious look finds a real and sizable loss.

Ahrefs sharpened the picture by asking what happens to the page that used to win. Looking at top-ranking pages, Ahrefs first measured the page in first position losing about 34.5% of its clicks when an overview appeared above it, and a later, larger analysis put the drop closer to 58%. That figure matters more than the averages, because the first result was the prize the whole search industry was built to capture. Losing more than half of its traffic to a box that sits above it changes the math on what a number one ranking is even worth.

Not every page felt it the same way

The averages hide the real story, which is that the loss was not spread evenly. Informational and commodity queries took the worst of it. If someone wanted to know how many ounces are in a cup, what year a war ended, or how to convert a unit, the overview handed them the answer and the visit never happened. Those were never high-value clicks. They rarely led to a sale, a signup, or a return visit. But there were a great many of them, and for sites whose traffic charts were built on that kind of question, the floor moved. A page that existed to settle a fact in one sentence has very little defense against a system that can settle the same fact in one sentence and never make the reader leave.

Commercial and high-consideration queries held up far better, and in some cases did more than hold. When a question carries money or risk behind it, which laptop to buy, which vendor to trust, how to handle a health problem, people still want to read, compare, and check who is saying it. A single summarized line does not close that loop, because the searcher knows the stakes are too high to take one sentence on faith. Those visits largely survived. In some cases the overview even sent more qualified traffic, because being named or cited inside the answer is its own kind of referral, and a reader who clicks after seeing a source vouched for tends to arrive further along and more ready to act.

The mechanism behind the split is not mysterious. AI answers are best at the questions with a single correct response and worst at the questions where the right answer is "it depends on you." The first group is exactly the commodity click. The second is exactly the considered one. So the technology did not flatten organic traffic so much as sort it, taking the cheap clicks and leaving most of the valuable ones in place.

This sorting matters for how a loss should be read. A site that watched its sessions fall by a third may have lost almost nothing of value, if what left was the unit-conversion and definition traffic that never converted. Another site with a smaller headline drop may be in real trouble, if the visits that vanished were the ones near a purchase. The aggregate number on a dashboard cannot tell those two cases apart. Only a look at which queries moved can, which is why the teams that reacted well spent their time segmenting the loss rather than mourning the total.

What AI answers actually cite

If the considered click survives, the next question is what earns its way into the answer, because that is where the surviving traffic now flows from. The evidence here is more specific than most teams expect. In an Evertune analysis of around 25,000 of the most cited pages across the major models, published by Search Engine Land, ranked best-of lists were the single most cited content format by a wide margin. Depending on the model, listicles accounted for somewhere between roughly half and two-thirds of the most cited URLs, with explainer guides and comparison pages behind them and single-product reviews further down.

The reason is structural rather than mystical. A page titled with the exact question a reader asked, followed by a set of named options each described in the same shape, is already built like an answer. A model can lift it almost without editing, which makes it cheap to cite and easy to trust. Loose, unstructured prose forces the model to do the synthesis itself, and given a choice it reaches for the page that already did that work. The lesson is not to chase a format for its own sake. It is that clarity and structure are now read by machines as well as people, and the machines reward the page that made itself easy to use.

Where the sources come from

The other surprise is how often AI answers lean on community discussion. Reddit and forum threads rank among the most cited sources across most studies, and Semrush reported that Reddit at one point accounted for around 40% of the citations it tracked. That figure should be held loosely. It swings widely by model and by month, and other trackers have put Reddit well below that mark, so the precise share is unstable in a way the underlying pattern is not.

The pattern is that models reach for forums because the threads carry unguarded opinion from people with no stake in the sale. That is the kind of source a polished vendor page cannot manufacture. When a model is asked which tool is actually worth buying, a real argument among real users reads as more credible than any marketing page, and the model treats it that way. For a brand, this means part of the conversation that decides whether you get cited is happening somewhere you do not own and cannot edit, which is uncomfortable and also true.

It also reframes what the citation data and the click data say together. The same shift that took the commodity click also rewired where the surviving clicks originate. A reader no longer arrives because you ranked first for a keyword. They arrive because a model read a best-of list, a comparison, and a forum thread, weighed them, and decided you belonged in the answer. The story and the numbers point the same way. Visibility now runs through being selected by a system that has already read the field, and the inputs to that selection are structure, usefulness, and what other people say about you when you are not in the room.

What it adds up to

The takeaway is not that search is dead. It is that the cheap, commodity click is dying and the considered one is not, and the two need to be managed as separate problems.

• Stop counting on commodity traffic. If your numbers came from answering questions a single sentence could settle, that traffic is leaving and it is not coming back. Planning around it only delays the reckoning
• Build for the considered click. The visits that survive belong to people making a real decision, so the job shifts from ranking a page to being the source the answer trusts and quotes
• Write in the shape that gets cited. Ranked lists, clear comparisons, and structured explainers earn citations because a model can use them directly. A strong best-of list is worth more now than another thin blog post
• Watch the rooms you do not own. Community discussion feeds these answers, so what gets said about you on places like Reddit is now part of your visibility, not separate from it
• Measure being named, not only ranking. The old scoreboard tracked position. The new one has to track whether the answer cites you at all, because that is where the surviving traffic starts

None of this is a trick, and most of it is the advice that always held, with the bar raised and the audience now made partly of machines. Be the most useful and most trusted source on your topic. The teams that accept the sort, defend the clicks worth defending, and let the rest go will be fine. The ones still chasing the volume that left will keep watching a number fall and wondering why.