The measured drop is 15% to 8%, on 68,879 real searches
Pew Research Center watched the actual browsing of 900 US adults for the whole of March 2025, which came to 68,879 Google searches. It is one of the few numbers in this space taken from what people did rather than what a tool estimated.
In that data, 18% of searches returned an AI summary at the top. When one appeared, people clicked through to a normal result 8% of the time. When one did not, they clicked 15% of the time, close to twice as often. Clicking a link inside the summary itself happened in 1% of visits.
Read that last figure slowly. The summary answers the question, and the sources it stands on are almost never opened. The page that fed the answer does the work and does not get the visit.
One answer is taking the place of a page of links
This is not only a Google story. OpenAI has said ChatGPT passed 800 million weekly active users during 2025, and a large share of those sessions are people asking for a recommendation, not a list of ten blue links to sort out for themselves.
When a buyer asks an engine for the best tool in a category, it does not hand back a ranked page. It names a handful of companies in a sentence or two and stops. The buyer reads the names, not the citations, and moves on. Whether your name sits in that sentence is now the whole game.
The old search game rewarded whoever had ground rank for a decade. This one has almost no incumbents, because the shortlist is being written now, by machines, for buyers who never see a link.
Your page can look perfect to you and be absent from the answer
The awkward part is that none of this shows up in your own analytics as a problem. The page still ranks, still loads, still looks right. It is simply not the page the model reached for when it built its answer, so it never enters the sentence the buyer reads.
Models do not invent their shortlists. They assemble them from a small set of pages they can read and trust. In the category we tested, listicles were 44% of what ChatGPT cited for a best of question, review sites 24%, vendor pages 18%. That figure is ours and it is scoped to what we sampled, not a law of the web, but the shape holds: a short, knowable set of sources decides who gets named.
The shortlist is worth entering because it holds
A list that reshuffled at random every day would not be worth chasing. The reassuring finding is that it mostly does not. In our own sampling the same five to seven names held across almost every phrasing of the same question. Ask it again tomorrow and some of the list churns, between 9 and 27% day over day depending on the engine, but the core is stable enough to be worth entering and stable enough that entering it holds.
That stability is the whole argument for treating this as a field to enter rather than a loss to mourn. The slot is open, it is cheaper to take now than it will be later, and once you are in the sentence you tend to stay.
The advice we tell people to skip
Some of the credibility here comes from what we decline to sell. The most recommended item in this space is publishing an llms.txt file. None of the engines we watch documents llms.txt as a ranking signal. Selling it as the fix would be theatre, so we tell people not to, including when saying so costs us a sale.
The cheapest real gap is usually the listicles, because most of them take submissions. If a page that already gets cited for your category will consider adding you, that is a far shorter path into the answer than rewriting your whole site.
Where this could be wrong
Google has disputed the Pew figures, arguing the study window overlapped with unrelated ranking tests. The study is also one month, one country, and a snapshot of results that change from week to week. Independent estimates of the click drop land wider, from roughly a third to about a half, depending on who measured and when.
Our own numbers carry the same caution. The 44% is a single measurement in the categories we sampled, not a settled constant, and we say so on the page they came from. The honest reading of all of it is a clear direction, one answer in place of ten links, with real uncertainty on the exact size of the effect.
None of this is a reason to panic, and none of it is a reason to buy anything today. It is a reason to check one thing you have probably never checked: whether the engines name you when a buyer asks. You cannot fix a sentence you have not read.