Google AI Mode Cut Web Clicks by 18.8 Points. Now It Is the Homepage Button

Forcing Google users into AI Mode cut their clicks to external websites by 18.8 percentage points. That is the headline result of a preregistered field experiment on 1,100 US Chrome users, posted to arXiv on 18 August 2026. Google AI Mode now sits on the homepage as the primary button for many users, and the company says it passed 1 billion monthly users. The experiment is the only causal estimate of what that combination does to publisher traffic.

What changed on the Google homepage, and when

The dates in circulation are muddled, so here is the sequence as the primary sources give it. Google added an AI Mode button to the right of the search box on its desktop homepage on 29 July 2025, announced by Robby Stein, Google Search's vice president of product. That button has been there for over a year.

On 11 August 2026 the classic Search button disappeared for some signed-out desktop users, replaced by three shortcuts labelled Create images, Ask about files and Brainstorm. Stein called it on X "a small test on desktop to help people find new things they can do with Search" with "no impact to how the core search box works".

On 29 September 2026, TIME reported that Google "has placed an AI Mode button on the Google Search homepage for many users". The photo caption states the homepage "now has an AI Mode button instead of a Search button". That is the change this post is about. Google has not published a rollout note for it, so the exact date and the share of users who see it come from press reporting, not from Google.

The button matters more than the feature. AI Mode has existed since May 2025. What changed is the default path. A user who types a query and reaches for the most prominent button now gets a conversational answer instead of ten links. The experiment below shows what happens to clicks when that path becomes the only one, and the homepage change moves the real world a measurable step toward that condition.

Google gives one number and withholds the other. Sundar Pichai told the I/O 2026 audience on 19 May that AI Mode "in just a year" had "surpassed 1 billion monthly active users," and that AI Overviews had passed 2.5 billion. Elizabeth Reid, who runs Search, wrote the same day that AI Mode queries were "more than doubling every quarter since launch". Neither post gives a figure for clicks from AI Mode to external sites. The only causal number on that question comes from outside Google.

How the AI Mode experiment actually worked

The paper is "AI in Search Reduces Publisher Referrals Without Improving User Experience," by Stephanie Wang and Danaé Metaxa at the University of Pennsylvania. Their co-authors are Jeffrey Gleason, Yakov Bart and Christo Wilson at Northeastern. It is a preregistered field experiment, which means the hypotheses, sample size and outcomes were filed before the data came in. That one fact separates it from every vendor click study published this year.

Participants were recruited on Prolific and through a Northeastern work-study programme between 17 and 19 March 2026. They had to be US-based adults using Chrome as their primary browser and Google as their primary search engine. 1,444 enrolled. 1,100 made at least one search during the treatment week and form the behavioural sample. Prolific participants were paid $10 in total.

The three arms

Each participant installed a Chrome extension and spent 3 days on unmodified Google so the team could record a baseline. Then the extension randomly assigned them, with equal probability, to one of three conditions for 7 days.

The extension logged every Google results page, every AI Mode response and every outbound click. Click-through rate was defined per user as external clicks from Google divided by Google searches, excluding clicks back to google.com and excluding vertical searches. Each conversational turn in AI Mode counted as a separate search, a choice that matters later.

What was preregistered

Three primary hypotheses covered trust, click-through rate and sessions per day. Four secondary hypotheses covered satisfaction, substitution to rival engines and related measures. The team ran a 72-person pilot to size the study and corrected for multiple comparisons with false discovery rate adjustment. The study was powered to detect a 7-point change in click-through rate, and it found one nearly 3 times that size.

What the 18.8-point AI Mode click drop means

The primary estimate is intention-to-treat. Everyone assigned to AI Mode is counted, including the 5.3% of their searches that slipped through to standard results. The point estimate is a fall of 18.8 percentage points in external click-through rate, with a confidence interval from 15.3 to 22.2 and a p-value below 0.001. The treatment-on-treated version, which adjusts for the users who got the full dose, comes out at 19.9 points.

A percentage-point change needs a baseline to become a relative one, and the paper does not state the control arm's click-through rate directly. It does give baseline averages across all participants: 9.2 searches a day and 4.1 external clicks a day. Dividing one by the other gives roughly 0.45 clicks per search. That is my arithmetic from the paper's figures, not a figure the paper reports, and it matches the 16% minimum detectable effect the authors describe for a 0.07 change.

On that baseline, 18.8 points is a fall of about 42% in clicks per search. My position: that number, not the 18.8, is the one to put in front of your board, with the derivation attached.

It was not only clicks. The AI Mode arm ran 0.92 fewer Google sessions a day against a baseline near 4.0. Clicks a day fell by 2.27. Trust dropped 0.34 on a 7-point scale, and satisfaction, usefulness and sense of agency all fell by more than half a standard deviation. In open-ended answers, 17.6% of the AI Mode group raised loss of control and 15.3% said it was hard to get to a specific website.

Removing AI did not simply reverse it

The No AI arm is the harder half of the paper. Among compliers, hiding AI Overviews raised click-through by 8.8 points. But the plain intention-to-treat estimate was 2.7 points and not significant. Google changed the AI Overview HTML during the experiment, and the extension went from hiding about 90% of Overviews on day 1 to close to 0% by day 10. The 8.8 is a local estimate that leans on an assumption the authors themselves flag.

Who lost the clicks under AI Mode

The paper breaks clicks out for four destinations only, chosen because they were the only ones with at least 5% of baseline clicks in the pilot. All four fell.

The ad figure is the one to read carefully. It is large because AI Mode carried no ads during the March 2026 test window, so the paper is measuring the absence of a surface rather than user behaviour toward it. That will change when Google monetises AI Mode, and the authors say so.

The category that is missing

There is no row for commercial sites, software vendors or comparison pages. The experiment recorded general search tasks and did not isolate high-intent product queries. So anyone who tells you the paper proves what AI Mode does to B2B pipeline is reading something that is not there. What it proves is the direction and rough size of the effect on clicks out of Google as a whole.

Reddit's number is a warning for vendors. Reddit took the largest measured hit outside ads. It is also where a large share of software buying conversations happen before anyone types a vendor name. If AI Mode absorbs the Reddit click, it is absorbing the step in the journey where your product gets compared, which is a step most B2B attribution never saw in the first place. The mechanics of how that community layer feeds AI answers are covered in the analysis of community mentions as an AI visibility input.

Building a B2B traffic model from the result

The experiment forced 94.7% of searches into AI Mode. Nobody's buyers are there yet. Before assignment, AI Mode handled 0.6% of participants' searches, and the authors are explicit that the AI Mode arm "reflects a forced, full-adoption scenario rather than the gradual, opt-in uptake" users experience organically.

So the useful model is linear in adoption. Suppose a share A of the searches that matter to you move into AI Mode, and each of those searches loses about 42% of its external clicks. Your expected organic click loss is A multiplied by 42%. Everything else held equal, which it will not be.

Three inputs are yours to measure. First, the adoption share for your buyers, not the global one. A 1 billion user headline tells you nothing about how often a head of procurement in your segment uses AI Mode for a vendor query. Second, your current share of the clicks on those queries, which Search Console gives you. Third, the share of your pipeline that originates from non-branded organic, which your CRM gives you if the attribution was set up honestly.

What the model ignores on purpose

It ignores the session effect. The AI Mode arm also ran fewer Google sessions, which would compound the click loss, but some of those sessions moved to rival engines where your page may still rank. It ignores the possibility that the clicks AI Mode does send convert better, a claim Google has made and nobody outside Google has verified. And it ignores ads, because the test window had none.

Three AI Mode adoption scenarios for a B2B site

The table applies the model to a site that gets 10,000 non-branded organic clicks a month from Google. Pick your own number and scale it. The relative loss per adopted search is 42%, derived above from the paper's 18.8-point drop against its 0.45 baseline.

The adoption shares in the first three rows are chosen as round scenario points, not measured. The relative loss is derived from one 7-day experiment on a sample that skews young and educated. Replace the 42% with your own figure the moment you have one.

Which scenario the homepage change points at

The homepage button does not force anyone into AI Mode. It changes the default for people who reach for the obvious button, which is a different mechanism from the experiment's network-level redirect. My reading is that it moves general consumer queries faster than it moves vendor-evaluation queries, because a buyer with a shortlist already knows which site they want. That argues for the 10% to 30% band for most B2B sites over the next two quarters, and I would plan against 30%.

The same paper carries a counter-signal. 11.2 percentage points more of the AI Mode group searched on Bing, DuckDuckGo or Yahoo during the week, and stated intention to switch to Bing rose 1.25 points on a 7-point scale. If a forced default pushes a measurable slice of users to a rival engine where your pages rank in a conventional list, part of the lost Google click comes back somewhere else. The paper did not measure whether those rival-engine searches produced clicks, so this is a possibility, not an offset.

Where this argument is weakest

A post that models revenue loss from one preprint owes you the holes. There are four, and the authors name most of them.

Seven days of a forced default is not a year of a chosen one

Participants were pushed into an interface that handled 0.6% of their searches before the study. Some of the click drop could be unfamiliarity rather than AI Mode's design, and the paper's own survey shows users resented the loss of control. People who choose AI Mode a year from now will be different people with different habits, and the effect on them could be smaller or larger. The experiment cannot say which.

The sample is not your buyers. 75% of participants were under 45, 87% had some college education and the group leaned left politically. Prolific workers are not procurement directors. The direction of the effect is unlikely to flip in a different sample, but the size could move a long way, and the model above inherits every point of that uncertainty.

The baseline click-through rate is my derivation

The 42% relative figure that drives the scenario table rests on dividing two reported averages, 4.1 clicks a day by 9.2 searches a day. That is a ratio of means, not the mean of the per-user ratio the paper actually uses as its outcome. If the true control-arm baseline is higher, the relative loss is smaller. I have shown the working so you can disagree with it.

Google's case has a real number behind it

Google says people who use its AI features search more, and that queries hit an all-time high in the first quarter of 2026. The experiment measured sessions per day falling under forced AI Mode, which contradicts that for its sample and its week. Both can be true if voluntary AI Mode users differ from forced ones. Until Google publishes click data by surface, that disagreement stays open, and the honest position is that the experiment measures the forced case and Google measures the voluntary one.

What to change in your organic plan

Three changes, in order of cost.

Split your reporting before the drop arrives. The 30% scenario looks like a ranking loss on a standard organic dashboard, because clicks fall while positions hold. If branded and non-branded queries, impressions and clicks are reported as one line, you will spend a quarter chasing a ranking problem that does not exist. The replacement reporting structure is laid out in the piece on what to report once organic sessions stop being the scoreboard. Search Console's generative AI performance report should be a tab in it from this month.

Measure what AI Mode says about you, not only what it sends. Being cited inside an AI Mode answer and being clicked from it are different events. The analysis of how AI citations broke away from Google rankings shows the overlap is already low, so a page that ranks third may be absent from the answer that replaced the click. A monthly citation panel of your buyers' real questions costs an afternoon and is the only series that will explain the gap.

Put the referral quality claim to the test. If AI Mode sends fewer visitors, the only consolation is if they convert better. That claim exists and has supporters, and the evidence behind it is weighed in the review of AI referral conversion multiples. Do not accept it on your behalf. Tag AI referrals separately, give them 90 days, and compare pipeline per visit against the organic baseline you already have.

Frequently asked questions

How to run your own version this month

You cannot randomise your buyers. You can measure the two sides of the model with what you already have.

Pull the 100 non-branded queries that drove the most clicks in the last 90 days from Search Console. Run each one through AI Mode and note whether your domain appears in the answer, and whether a link to it is visible without expanding anything. That gives you a citation rate and a visible-link rate for the queries that actually pay you. Repeat it on the same date next month, and the slope of that line is your own adoption-adjusted version of the 18.8-point number.

Related reading