Yesterday I published a piece about my own collapsed Google traffic, and about how I blamed AI when it was really my own domain migration. Decent story, but it was one shop and one graph. Losing traffic to AI is the explanation everyone reaches for right now, and I reached for it too. I did note myself that AI search behaviour genuinely exists, but that in my case it played a supporting role and not the lead. At that moment that was a gut-feel conclusion with exactly one data point underneath it: mine.
And then a report landed in my inbox that extends that conclusion to tens of millions of users. Timing you cannot plan, so I might as well write about it straight away.
It's the State of Search report on Q2 2026 by Datos, these days part of Semrush. Not a survey where people say what they think they do, but clickstream: the actual click behaviour of tens of millions of desktop users in the US, the EU and the UK, across fifteen months. The commentary in the report comes from Rand Fishkin, founder of SparkToro and before that of Moz. I've followed him on LinkedIn for years, and he belongs to the handful of people in this industry who consistently share data instead of opinions. His remarks alongside these numbers are, as far as I'm concerned, the best part of the report, so you'll come across him a few times below.
The outcome is fairly uncomfortable for the story we all keep repeating.
Losing traffic to AI: far smaller than you think
As a percentage of all desktop activity in Q2 2026:
| Category | US | EU and UK |
|---|---|---|
| Traditional search engines | 10.52% | 11.63% |
| AI tools (ChatGPT, Gemini, Claude, all combined) | 1.83% | 1.91% |
| E-commerce | 2.79% | 1.96% |
| Google AI Mode | 0.12 to 0.13% | 0.29% |
Search is roughly six times as large as all AI tools put together. And it grew this quarter. Not spectacularly, but it rose.
AI grew faster, granted: from 1.37 to 1.83 per cent in the US in a year. But in June something happened that hadn't happened in years. The AI curve dipped. In both regions, after a peak in May. Fishkin is cautious about it: one month is not a trend, but this is the biggest drop in years. It could be a hiccup. It could be a plateau. It could also be that Google has by now crammed so much AI into its own results that you no longer need to go anywhere.
That dip sits in the red bars of the monthly graph. AI tools climbed almost uninterrupted for fifteen months, from 0.64 per cent in April 2025 to 1.10 per cent in May 2026. And then June: 1.05 per cent. The first step back in over a year.
Then look at the blue bars underneath. Traditional search has been fluctuating between 9.4 and 10.8 per cent for fifteen months, and June 2026 (9.83 per cent) sits at virtually the same level as June 2025 (9.78 per cent). There is no erosion. There is a line running dead straight while everyone says it's collapsing.
Amusing detail for anyone who thinks Europe is lagging behind: Europe uses more AI than the US. And Google AI Mode, which launched later here, sits at more than double the American level in Europe. That curve climbs steeply and at the same time sits on an axis that stops at 0.40 per cent. Which is precisely the point: it's growing fast and it's still nothing.
A warning about every market share figure you'll come across this year: reach is not usage. Google reported that AI Mode passed a billion users per month. Sounds overwhelming. In this measurement AI Mode sits at 0.12 per cent of desktop activity. Both numbers are true, they simply aren't measuring the same thing. Trying something once is not a habit, and a habit is what determines your traffic.
This is the number that actually matters
While everyone stares at ChatGPT, something else is happening. In Europe, clicks after a Google search going to Google-owned destinations rose from 11.7 per cent in April 2025 to 20.4 per cent in June 2026. Almost nine percentage points added in a year.
One in five clicks after a search never leaves the building.
Over the same period, organic click-through fell from 46 to 40.7 per cent. Fishkin calls that the number that bothers him most, and he points out that this is desktop, where there should be more clicking than on mobile.
Zero-click, which the whole industry keeps waving around, went from 24.7 to 25.3 per cent in Q2. Higher than last quarter, lower than last year, and measured across fifteen months without a clear direction. If you come across the "68 per cent zero-click" headline somewhere: that's a different measurement with different definitions, not a refutation of this one. But it does show how elastic that number is, and how easily you can build a panic story around it.
The summary in one sentence: you're not losing your click to a chatbot, you're losing it to the search results page itself.
People search differently, and that's your opening
This was the part I found most interesting, because you can do something with it tomorrow.
Navigational search is collapsing across every search engine. At Google in Europe from 28.69 to 20.57 per cent, in the US from 22.77 to 16.43. At Bing from 32.81 to 20.63. At DuckDuckGo almost halved. People are typing a brand name into the search bar to get somewhere less and less.
What replaces it is informational and orientational search. At Google in Europe, informational went from 55.39 to 61.31 per cent and commercial research from 14.99 to 17.47. Everywhere, every search engine, both regions moving in the same direction. And the queries are getting longer: searches of six and seven words have grown four quarters in a row.
Purchase intent, meanwhile, is stuck at 0.65 per cent of all searches at Google in Europe. Six hundred and sixty-five thousandths. That's no news to anyone who already knew, but it is a solid reality check if you've built your entire content plan around "buy" terms.
What this means for a webshop: the volume is shifting to precisely those questions that come before the purchase. Comparisons, suitability, maintenance, what suits whom. Longer questions, more specific, with more context. That is not the domain of your product page. That is the domain of content that actually knows something.
The annoying part is that this is also exactly the kind of question an AI summary is happy to handle itself. So you don't win that volume by throwing thin texts at it. You win it by being the source that gets cited, and that only works with something a language model can't invent for itself. Your own measurements, your own numbers, your own mistakes. Why I keep hammering on that is in SEO for webshops: where to start in 2026.
The AI market itself is shifting
For anyone wondering which assistants they actually want to be mentioned in:
- ChatGPT is still the largest, but slipped in Europe from 44.8 per cent user share in March to 41.9 per cent in June.
- Gemini stood at 22.9 per cent in Europe at the end of Q2. Google is pushing it in everywhere.
- Claude grew from 9.6 to 12.6 per cent in Europe. Fishkin thinks ChatGPT is becoming the consumer counter and that professionals are moving to Claude and Gemini.
- DeepSeek passed Perplexity for the first time.
And then the pattern that reminded me most of Google: referrals are disappearing out of AI tools. Wikipedia, NIH and ResearchGate all dropped out of the European top 15 of destinations people go to from AI. Social platforms fell across the board. What's rising is other AI tools: people hop from ChatGPT to Gemini to Claude.
Google keeps users inside Google. AI assistants keep users inside AI. Everyone is building the same fence, just around a different garden.
One number almost nobody is acting on
Reddit grew in Europe from 29.4 per cent of desktop users in April 2025 to 41.6 per cent in June 2026. Four in ten.
Fishkin adds that he's baffled that so few marketers are doing anything there, while they knock themselves out on LinkedIn. I agree with him and I don't do enough of it myself either. Duly noted.
What I'm taking away from this
Five things, and they're expanded further down this page. Briefly: look before you blame AI, account for the click Google takes for itself, write for the orientational question instead of the buying term, become machine-readable while it's still cheap, and build the traffic that is yours.
That fourth point, becoming machine-readable with an open product feed and an llms.txt, I worked out in The future of e-commerce: I asked AI, including the addresses where you can inspect our own feed. And if a migration or platform switch is on the calendar, that's the moment your visibility moves hardest: the full playbook is in Switching to a new webshop platform.
The small print, because it belongs here
This report measures desktop only. On mobile, where most search traffic sits, it looks different, and probably worse for the click. Threads, for example, doesn't appear in it at all, while it reported 500 million active users in June, simply because it's an app.
It's also a panel and not a census. And Datos belongs to Semrush, which has an interest in you continuing to take search engines seriously. That doesn't make the numbers wrong, but I'd rather say it myself than have you look it up.
Source
State of Search Q2 2026, Datos (A Semrush Company). Clickstream from desktop panels in the US, the EU and the UK, April 2025 through June 2026, with commentary by Rand Fishkin of SparkToro. The graphs in this article come from that report and are reproduced unchanged with attribution.
Why you can't download the report here. You can, just not from me. Datos gives it away for free in exchange for your email address, and that form is the only thing they get back for research they invest heavily in. If I put the file here, I'd be using their work while demolishing their side of the trade. That doesn't strike me as a way to treat a source you then build an entire article on. There's a practical point on top of that: the download link I received is personal and contains tracking codes tied to my own sign-up, so I couldn't even share it without everyone who clicks it being counted as me. Citing two graphs with attribution is fine. Passing the file along is not.
If you want all the graphs and the full analysis, get it straight from Datos. It costs you an email address and it's worth it. I wrote a blog off the back of it immediately, so I'm not objective.