Google removed a click. It also removed the comparison step.

The change itself is small enough to miss. For certain queries, the AI Overview no longer waits for a click to become the fuller answer. It expands on its own, brings the follow-up box with it, and pushes the traditional results further down the page. Chris Long spotted it in the wild in late August, Barry Schwartz confirmed it, and Google described it as dynamic expansion for topics where its systems determine it is most useful.

Treated as a layout change, it is a minor story about a button. Treated as a behavior change, which is what the research says it is, it is one of the more consequential things to happen to how people evaluate answers this year.

What Google changed, and when

The mechanics, precisely. Previously an AI Overview presented a summary with a Show more control, and a user who wanted the longer conversational answer clicked it and landed in AI Mode. Now, for a subset of queries, the expansion happens without the click. The Ask anything follow-up box appears by default. Google told Search Engine Roundtable that for some queries, AI Overviews may dynamically expand for topics where its systems determine it is most useful for people, and noted that users who have already started scrolling will not have the page expand underneath them.

That last detail is a courtesy to reading position, and it is also a tell. Google is treating the expansion as something that happens before engagement begins, which means the user meets the expanded answer as their first impression of the result page rather than as something they asked for.

Two things follow immediately. The organic results move further down. And the user arrives in a different interaction pattern than the one they would have been in a week earlier for the same query. The first is the consequence everybody is discussing. The second is the one with the data behind it.

AI Mode vs AI Overviews: two surfaces, opposite behavior

Clickstream Solutions analyzed roughly 846,000 US Google search sessions from February and March 2026, drawn from anonymized clickstream data supplied by Surfer SEO, alongside a series of smaller instrumented studies. The finding worth carrying around is that the two surfaces do not produce a difference of degree. They produce opposite postures.

The AI Overviews column describes someone shopping. Cursor scattered wide, scrolling up and back down, holding still while reading, checking one thing against another. Among users who reversed direction at all, the median spent 47.5% of their scrolling going backward when an AI Overview was present, against 27% without one. That is what deliberation looks like in mouse data.

The AI Mode column describes someone being handed a conclusion. In an April 2026 study of 185 high-stakes purchase tasks, users accepted the shortlist the AI produced in 88% of cases, selected the top-ranked item in 74%, and clicked nothing at all in 64%. High-stakes, note. Not casual lookups. The kind of decision where you would expect a person to open three tabs.

“One surface makes people compare. The other makes them accept. Google just made the second one the default arrival state for a growing share of queries, without anybody choosing it.”

The 64% who click nothing

It is tempting to read 64% as a traffic story and stop there. It is a traffic story, and the zero-click math has been beaten to death elsewhere. The more useful reading is about what the remaining 36% are doing and why.

Look at 88% against 74%. Nearly nine in ten took the shortlist as the universe of options, and three in four then took the first thing on it. The competitive question is no longer only whether you get cited. It is whether you are on the shortlist and where you sit on it, because a meaningful majority of users are not going to interrogate the ordering.

That is a harsher ranking environment than ten blue links ever was, and it is harsher in a specific way: there is no second page, no scroll to position eight, no long tail of people who dig. Being fourth on an AI Mode shortlist is closer to being invisible than being fourth in organic results ever was. It also raises the value of the sources that shape those shortlists in the first place, which is the argument underneath building citation presence without a licensing deal.

Why the comparison step was doing work for you

Here is the part that is easy to miss while counting lost clicks. The comparison behavior AI Overviews produced was not just traffic. It was a chance to be evaluated.

A user scrolling back and forth across a results page, cursor wandering, checking one summary against another, is a user who can still be persuaded by something other than rank position. A better title. A more specific claim. A source they recognize. A review count. That user is running their own comparison, and a brand that is not the top-ranked answer can still win them.

The clickstream data caught something else in that behavior worth naming. When AI Overviews are present, query intent stops predicting time on page. Without one, navigational and local searchers behaved very differently from each other, a 20 point spread in who was still active 21 seconds in. With an AI Overview present, every intent type clustered together, between roughly 42% and 48%. Even people searching a brand name by name were behaving like researchers rather than like people going somewhere they already knew.

The report's own framing of that, that brand recall is no longer enough on its own, is worth taking seriously for anyone whose navigational queries have felt oddly soft this year. Users who typed your name are now reading a summary about you before deciding whether to proceed. Auto-expansion into AI Mode does not restore the old behavior. It replaces the deliberation with acceptance of whatever the summary said, which is fine if the summary is right about you and expensive if it is not.

What AI Mode vs AI Overviews means for the way you write a page

Three shifts follow, and none of them are new inventions so much as sharpened priorities.

The through-line is that hedged, incomplete, keyword-shaped pages degrade badly here. They degraded in AI Overviews too, but a comparing user could still rescue them by clicking through and finding the real answer. An accepting user never does. This is the same direction we tracked in the format shift running through AI search results, now with a stronger forcing function behind it.

What to change this quarter

Start by finding out which of your queries are affected, because the expansion is query-dependent and nobody outside Google has the list. Run your top 50 commercial queries manually, note which ones now expand without a click, and treat that as your priority set. It is an hour of work and it turns a general trend into a specific list of pages.

For those pages, audit for answer completeness first. Can a reader get the actual answer, with the specifics, without leaving the page? If the page is structured to withhold, restructure it. Then check whether the comparative attributes a shortlist needs are present as text rather than implied by a graphic or a table image, which is the most common version of this failure in content programs built before extraction mattered.

Then adjust what you report. A page that loses clicks while gaining shortlist inclusion is winning and will look like it is losing in any dashboard built on sessions. Pair the traffic number with a citation or inclusion measure for the same query set, or the quarterly review will conclude the wrong thing with total confidence. The carousel and citation slot mechanics we mapped in AI Mode are a reasonable place to start on what to count.

There is a sequencing point worth making explicit, because teams tend to get it backwards. The instinct after a change like this is to rewrite everything at once, which produces a quarter of undifferentiated work and no way to attribute the result. Do the query audit first, restructure the affected pages second, and hold the rest of the library steady long enough to see whether the restructured pages behave differently from the untouched ones. That comparison is the only evidence you will get that the work is doing what you think it is, and it evaporates the moment you change everything simultaneously.

One thing not to do: treat this as a reason to write for the model instead of the reader. The behavior that changed is the user's, not the machine's. People are accepting summaries because summaries got good enough to accept, and the response that works is being the source whose specifics survive summarization intact. That is the same job as before. The margin for a vague page just got a lot thinner, and the click that used to bail you out is gone for a growing share of queries.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.