Video via the New York Times
A New York Times article this month cited one of the AI Mode studies I led the research for on behalf of Kevin Indig and Growth Memo (gifted link). The finding: roughly 75% of AI Mode sessions were zero-click, never leading users out to the Web. Put the other way, about 25% of AI Mode tasks led users to visit the open web. A Google spokesperson pushed back. And if you want to track AI Mode traffic yourself to check either number, you can’t, which is the second half of this story.
If anything, 25% is generous.
Over the past year, I’ve led the UX research for three AI Mode studies, with Kevin leading the analysis and publication for two of them. Across those studies, involving more than 100 participants, the share of tasks that led people to external websites has consistently averaged below 25%.
In my most recent study for Garrett French and Citation Labs, only 22% of tasks resulted in participants leaving AI Mode. That study involved 48 participants completing deeper research tasks, exactly the kind of work where people would be expected to complement AI responses with independent websites.
If deeper research tasks skew the numbers at all, they skew toward more clicks to the open web, not fewer. The real-world number for everyday AI use is likely lower than what we measured.
Not just fewer clicks. No clicks.
The same Clickstream Solutions research done for Citation labs looked at 149 AI Mode tasks and found that participants clicked nothing 64% of the time. That figure is not “stayed on Google.” It is a true zero-click session: no cards, no panels, no links, no expand buttons. No follow-up queries either. They read the AI text and declared their finalists.
That’s not a search session.
I’ve done a plethora of these studies, most recently for @Profound, looking at ChatGPT. The same pattern shows up again and again. In that study, 92.8% of tasks ended without a meaningful click to the open web. The non-clicking user is not a Google policy choice. It is what happens when you put a confident synthesized answer where a list of options used to be.
What Google actually disputed
The pushback in the Times was that our study used a small sample and assigned participants specific tasks that don’t reflect normal behavior.
Assigned tasks have been the standard instrument in moderated usability research for three decades. Google’s own product teams run studies this way. You cannot observe how someone resolves a purchase decision without giving them a decision to resolve, and every large-scale alternative to assigned tasks has a worse problem, which I’ll get to below.
Google did not say our direction was wrong. Google did not offer a competing number. Google disputed how the measurement was taken, not what it found. That is the response of a party that has the data and has decided not to produce it.
The one number that would settle this
Session-level click-out rates for AI Mode, tracked separately from total Search. Not blended organic click volume across all surfaces, which is the figure Google keeps offering, and which can stay flat while AI Mode specifically sends almost no one anywhere.
Google has that number. It has never published it. Everything else in this argument is inference from the outside because the inside is closed.
So which is it?
Google’s 2024 promise was “Let Google do the searching for you.” In 2025 they said they care about the web ecosystem “perhaps more than any other company.” So which is it: they’re taking the information and keeping users on Google, or they’re still sending people to websites in similar numbers?
They claim it still sends a “relatively stable” number of clicks to the open web.
It’s worth remembering what the company said it was for. In the 2004 letter to shareholders before Google went public, Larry Page wrote that a well-functioning society should have abundant, free, unbiased access to high-quality information. Google honored that by being a gateway. It took your question and pushed you outward, to merchants and publishers and universities who had answered it.
The gateway is now the destination. That is not a critic’s framing. It is the stated product goal, written in Google’s own headline two years ago.
Why you can’t track AI Mode clicks
Google won’t release data of its own about user behavior in AI Mode, and third-party clickstream data can’t fill the gap because prompts are too detailed and too long-tail to aggregate.
Google itself supplied the reason at I/O: AI Mode queries run roughly three times longer than traditional keyword searches. Panel and clickstream vendors built their entire methodology on the assumption that many people type the same short query, which lets you aggregate to a query and measure what happened next. Conversational prompts are close to unique. There is no query to aggregate to.
So the tools you would normally reach for to track AI Mode traffic are structurally unable to see it. Rank tracking tells you whether you were cited, not whether anyone moved. Analytics tells you about the visits that happened, not the 64% of sessions that ended before a visit could exist.
Which leaves moderated think-aloud as the only method that currently observes a complete AI Mode session start to finish, including the moment someone reads a summary and quietly decides they’re done. Small samples are a real limitation of that method. They are also the price of being the only people looking.