Google AI Mode Cut External Clicks by 18.8 Points — A 1,100-User Field Experiment

Key takeaway: In a preregistered field experiment that randomly assigned 1,100 Google Search users to three conditions (Wang et al., arXiv:2608.18352, preprint), sending every search to AI Mode cut the share of searches leading to an external click by 18.8 percentage points. Hiding AI Overviews, by contrast, raised clicks by 8.8 points. Participants in the AI Mode condition also rated search satisfaction and trust lower, and more of them tried other search engines. Until now, most evidence that "AI search reduces clicks" came from observational data; this is experimental evidence that supports a causal reading. With Google starting to auto-expand AI Overviews in late August, here is what to measure besides clicks.

---

Three-line summary

• In the AI Mode-only condition, the external click share fell by 18.8 points; in the condition with AI Overviews hidden, it rose by 8.8 points. The drop was large for third-party sites such as Reddit, news outlets, and Wikipedia as well.
• Users did not get a better experience in exchange for fewer clicks. Satisfaction, usefulness, and trust all fell in the AI Mode condition, and the share of users who tried other search engines rose by 11.2 points.
• AI Mode is not yet Google's default search experience, but AI Overviews auto-expansion is moving the default page in that direction. If visits are your only KPI, you miss how your brand is treated inside the answer.

How the experiment was designed

The study changed the same people's real Google searches for a week and compared the results. Participants installed a browser extension that modified the Google Search page in real time according to their assigned condition.

Condition	Search page	Role
No AI Search	AI Overviews hidden; AI Mode entry redirected to regular search	What happens when AI features are removed
Current Search	Regular Google Search (AI Overviews and AI Mode available)	Baseline (control)
AI Mode Search	Every search redirected to AI Mode; regular tabs hidden	What happens if AI Mode becomes the default

Item	Detail
Sample	1,444 enrolled, 1,100 who searched during the treatment period, 956 who completed the post-survey
Location	United States
Timing	Recruited in March 2026; 3 baseline days followed by 7 treatment days
Preregistration	OSF
Authors and funding	Researchers at the University of Pennsylvania and Northeastern University; funded by the US National Science Foundation (NSF); no competing interests declared

Click-through rate was defined as "external clicks originating from Google ÷ Google searches." Clicks that stay on Google (such as follow-up searches) and vertical searches such as images, news, and shopping were excluded.

This design matters because of the strength of evidence. In the evidence grades this blog laid out in its analysis of a survey of 45 GEO studies, the highest tier is the randomized field experiment. Because users in the same period and conditions were assigned at random, differences between conditions can be read as effects of the AI features. And unlike studies written by authors at commercial measurement vendors using operational data, this one discloses its funding and conflicts of interest.

Result 1 — AI Mode cut external clicks by 18.8 points

Change in the external click share versus the control (current Google Search):

Condition	Change in external click share	95% confidence interval
AI Mode Search	-18.8 pp	-22.2 to -15.3
No AI Search	+8.8 pp	+2.3 to +15.3

Source: Wang et al., arXiv:2608.18352. Effects are estimates controlling for baseline values.

Interpretation. AI Mode cut the number of searches that lead to an external site by about 19 per 100 searches. It also means that today's Google Search, with AI Overviews alone, produces fewer external clicks than search without AI features (hiding AI Overviews added 8.8 points). Note that in the No AI Search condition, a mid-study change to Google's page structure meant only 51.1% of AI Overviews were actually hidden; the researchers also report estimates adjusted for this.

The drop was not limited to one type of site. In the AI Mode condition, the share of users who clicked each type of site fell as follows.

Click target	Change in share of users clicking (AI Mode vs control)
Ads	-42.7 pp (AI Mode showed no ads)
Reddit	-21.2 pp
News sites	-12.5 pp
Wikipedia	-9.9 pp

Reddit, news sites, and Wikipedia are third-party sources that AI answers cite often. Because even sites that AI answers cite get fewer clicks, being cited and being visited should be treated as separate metrics.

Result 2 — Fewer clicks, and no better user experience

The explanation "clicks fell because users got their answer right away" was not supported by this experiment.

Metric (AI Mode vs control)	Change
Satisfaction	-0.73 SD
Agency (feeling in control of the search)	-0.66 SD
Usefulness	-0.59 SD
Perceived personalization and relevance	-0.42 SD
Trust in information on Google (7-point scale)	-0.34 points
Search sessions per day	-0.92
Minutes per session	+0.43
Share of users trying other engines (Bing, DuckDuckGo, Yahoo)	+11.2 pp

Participants in the AI Mode condition searched less often but stayed longer per session, and rated the results as less useful and less trustworthy. The No AI Search condition, by contrast, showed no significant difference from the control on the survey measures. Removing AI features did not make users rate search worse.

In open-ended responses from the AI Mode condition (309 participants), 33.6% were negative and 29.3% positive. Common complaints were loss of control (17.6%), difficulty navigating to specific sites (15.3%), and limited links or source diversity (13.4%); the most common praise was time saved (14.0%).

The researchers summarized their results this way:

"We show that removing AI Overviews and AI Mode increases click-through rates to publishers, while an AI Mode-only experience reduces click-through rates and erodes user experience and trust in information found on Google." — Wang et al., arXiv:2608.18352 (2026-08-18)

Why this experiment matters now — the default page is moving toward AI Mode

AI Mode is not yet Google's default search experience. As confirmed in this blog's fact-check, Google has said publicly that it has no plans to make AI Mode the default. At the start of the experiment, participants used AI Mode for only 0.6% of searches. The experiment's "AI Mode Search" condition is not reality; it models what happens if AI Mode becomes the default.

The default page, however, is moving in that direction. In late August 2026, Google changed AI Overviews for some queries to expand automatically without a "Show more" click, with the follow-up input box open from the start. A Google spokesperson described the change this way:

"For some queries, AI Overviews may dynamically expand for topics where our systems determine it's most useful for people." — Google spokesperson, via Search Engine Roundtable (2026-08-28)

When AI Overviews take up more of the page, regular results move further down. The default page is shifting to somewhere between the experiment's "Current Search" and "AI Mode Search" conditions.

Korean users are affected too. Google AI Mode expanded to five more languages, including Korean, in September 2025 (Search Engine Roundtable). This experiment, however, covered US users only, and Korea is a market where Naver holds a large share of search, so there is no basis for assuming the same effect size. The structure of Korea's search ecosystem is covered in our Korean AI search landscape.

Interpretation — "clicks fell" is now a causal finding, not just an observation

In July, this blog covered research on referral loss from ChatGPT search. That was an observation that only 5.2% of ChatGPT conversations produced an external click. Observational data struggle to answer the question "are people who use AI simply people who click less?" This experiment answers it by randomizing similar users, and provides causal evidence that AI answers reduce external clicks inside Google Search as well.

For brands, the result means three things.

1. Fewer visits do not mean less visibility. Even when your brand is mentioned and cited as a source inside an AI answer, users may finish their decision inside the answer without clicking. Visits alone do not capture that exposure.
2. Cited sources get fewer clicks too. Sites that AI cites often, such as Reddit and Wikipedia, also lost many clicks under AI Mode. The goal of earning citations shifts from clicks to being reflected in what the answer says.
3. Users leave when they do not like the answer. Complaints about difficulty reaching specific sites and the rise in other-engine use suggest that brands users know by name and navigate to directly may be better placed. This experiment did not measure branded searches separately, though, so this remains a hypothesis.

Putting it into practice — metrics to place beside the click KPI

Keep measuring clicks, but build a system that does not judge by clicks alone.

Metric you watch now	Metric to add	How to check
Google organic clicks	Impressions inside AI features	Search Console generative AI report (impressions, pages, countries, devices, dates; no clicks)
Search rank	Brand mention rate and own-domain citation rate in AI answers	Re-run the same questions and record mentions and citations
Post-visit conversion	Accuracy of how the answer describes your brand	Save answer text and check for factual errors and omissions

1. Read Search Console click declines alongside impressions. Google's Search Console generative AI report shows impressions inside AI Overviews and AI Mode but no click data. If impressions hold while clicks fall, you may be seeing the structure this experiment found. If impressions fall too, check whether you have dropped out of the AI answer's source selection.
2. Measure exposure inside the answer. Metrics that replace rank are covered in our AI visibility tracking metrics post. RanketAI's AI brand visibility analysis measures brand mentions and own-domain citations in ChatGPT, Gemini, and Perplexity answers. Answers on the Google AI Mode page are not measured, so if AI Mode answers matter to you, you can also pick a few key questions and record those answers yourself.
3. Keep the answer text. As clicks decline, the answer body becomes everything the user sees. Whether the answer describes your brand accurately can only be checked if you save the original text.

Five lines to hold when citing this study

The evidence is strong, but accurate citation requires stating the conditions.

1. It is a forced AI Mode condition. AI Mode accounted for 0.6% of searches before the experiment, and the researchers note that forced switching differs from gradual, voluntary adoption. Do not write "AI Mode cuts clicks by 18.8 points"; add the condition "if every search is switched to AI Mode."
2. The treatment lasted seven days. Long-term effects after users adapt to AI Mode were not measured.
3. The sample does not represent the US population. Participants were younger (75% under 45), more educated (87% college-educated), and politically skewed compared with the US population.
4. Treatment in the No AI Search condition was incomplete. A mid-study change to Google's page structure meant only 51.1% of AI Overviews were hidden. Read the +8.8 points with that in mind.
5. It is a preprint that has not been peer reviewed, and high-stakes searches such as health, legal, and financial queries were not analyzed separately. It is also not evidence about Korean-language search.

FAQ

Related reading

• AI Search Answers Without Referring: 9.4% Search Displacement, 5.2% Outbound Clicks — an observational study of referral loss from ChatGPT search.
• Did Google Kill the Blue Links? What Primary Sources Actually Confirm — whether AI Mode is the default, checked against primary sources.
• Google Search Console's New Generative AI Reports — the official view of impressions inside AI features.
• How to Track AI Search Visibility — Three Metrics That Replace Rank — metrics to use instead of clicks and rank.
• The 40% GEO Gain Has a Condition — Inside the 2026 Critical Survey of 45 Studies — criteria for judging the strength of research evidence.