Field experiment: Google AI Mode cut clicks to websites by 18.8 points

A preregistered field experiment with 1,100 US Google users found that forcing searches into Google's AI Mode cut click-through to external websites by 18.8 percentage points, while hiding AI Overviews raised click-through by 8.8 points. The study, by researchers at the University of Pennsylvania and Northeastern University, was posted to arXiv on 18 August 2026 and picked up by Search Engine Journal on 21 September.

It is one of the few causal studies of Google's AI search features. Most of what marketers know about AI Overviews and traffic comes from observational data, where many things change at once. Here, participants were randomly assigned to different versions of Google Search and tracked through a browser extension, so the differences between groups can be attributed to the AI features themselves, within the limits the authors set out.

How the experiment worked

The paper, "AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence", is by Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danaé Metaxa. Participants were recruited between 17 and 19 March 2026 through the research platform Prolific and a Northeastern work-study programme. All were US-based adults who used Chrome as their main browser and Google as their main search engine.

Everyone first had three days of normal Google Search, which gave a baseline. Then each person was randomly assigned for seven days to one of three conditions:

• No AI Search: AI Overviews were hidden, including those at the top of results, mid-page and inside People Also Ask, and AI Mode searches were redirected to regular search.
• Current Search: Google as normal, with AI Overviews where Google showed them and AI Mode available.
• AI Mode Search: every search was redirected into AI Mode.

Of 1,444 people enrolled, 1,100 made at least one search during the seven days and form the behavioural sample. 956 completed the follow-up survey. During the baseline, participants averaged 9.2 searches and 4.1 clicks a day, 36% of their searches triggered an AI Overview, and they used AI Mode for only 0.6% of searches.

The main results

The table lists the headline effects reported in the paper, each measured against the Current Search group.

AI Mode kept people inside Google longer per session while sending them to other websites less often and cutting how many sessions they started each day. The drop in daily sessions was larger among heavy Google users, those with more than four sessions a day, at about 2.01 fewer sessions.

Which sites lost clicks

AI Mode reduced the share of users who clicked through to news sites by 12.5 points, to Reddit by 21.2 points and to Wikipedia by 9.9 points. The authors say their estimates suggest the traffic effect reaches well beyond Wikipedia and news.

The share of users clicking on ads fell by 42.7 points in the AI Mode group. The paper is explicit about why: AI Mode did not show ads at the time of the experiment. That finding describes a gap that existed in March 2026, not a stable effect, because Google has since started testing ads inside AI Mode.

How people felt about AI Mode

Being assigned to AI Mode lowered perceived usefulness, satisfaction, sense of control and perceived relevance of answers. It also increased stated intention to switch to Bing, by 1.25 points on a seven-point scale. Hiding AI Overviews, by contrast, produced no detectable change in trust or search experience. In other words, taking AI Overviews away sent more traffic to websites without users reporting a worse experience.

When asked for their overall impression after a week, 33.6% of AI Mode participants were negative, 29.3% positive and 13.7% mixed. The most common complaints were feeling less in control (17.6%), difficulty reaching specific websites (15.3%) and too few links or sources (13.4%). The most common praise was time saved (14.0%).

The limits the authors set out

The paper is candid about its weaknesses, and they matter for how far the numbers travel.

• Hiding AI Overviews partly broke. The extension hid 90% of AI Overviews on day one, but a change Google made to AI Overview HTML during the study stopped it working, and success fell to zero. Overall, 51.1% were hidden, so the authors report adjusted estimates for that group.
• The sample is not representative. Participants skewed younger (75% under 45), more educated and more left-leaning than the US population.
• It ran for seven days. Longer-term adaptation to AI Mode was not measured.
• AI Mode use was forced. Baseline use was 0.6% of searches, so the AI Mode group reflects full adoption, not the gradual, opt-in use most people have.
• It is US-only and says nothing directly about other languages or markets.

The paper is a preprint on arXiv. It has not yet been through journal peer review, although the study design was preregistered before data collection.

How it fits with earlier evidence

The authors place their results next to earlier, observational work. They cite a study that found AI Overviews reduced English Wikipedia traffic by 15%, and say their own estimates suggest the effect generalises well beyond Wikipedia and news. That 15% figure has since been revised: a later working paper, covered by Relevant Audience on 13 September, put the Wikipedia effect at about 5%. The experiment does not settle the size of the effect for any one site. What it adds is causal direction, measured on the same people under controlled conditions.

The paper also points to a tension for Google. AI Mode is built to compete with conversational tools such as ChatGPT, and it keeps people on Google longer per session. Yet in this sample it pushed a meaningful share of users towards rival search engines. The authors argue Google will have to balance pushing AI Mode on users against the risk of alienating them.

What this means for Thai marketers

The study does not measure Thai users or Thai-language search, so the exact numbers should not be applied to Thailand. The direction of the results is still useful for planning. If Google moves more queries into AI Mode, sites should expect fewer clicks per search even where they are cited, and longer sessions that stay on Google.

Three practical points follow:

1. Measure visibility, not only clicks. If AI Mode cuts click-through, a page can be influential in answers while its traffic falls. Tracking whether your brand is cited in AI answers matters alongside Search Console clicks, which is the core of a GEO programme.
2. Make the click worth it. Participants complained that it was hard to reach the specific site they wanted. Pages that offer something an AI summary cannot give, such as tools, prices, booking and original data, give people a reason to click.
3. Watch ads in AI Mode. The 42.7-point ad click drop came from ads being absent. As Google tests ads in AI Mode, the paid side of this will change, and search budgets should be reviewed as those formats reach Thailand.

For teams deciding how much to invest in content built for AI answers, AI SEO work and traditional SEO now need to be measured with different metrics, because the same ranking can bring very different traffic depending on which Google interface the searcher sees.

FAQ

How much does Google AI Mode reduce clicks to websites?

In this experiment, forcing all searches into AI Mode reduced click-through to external websites by 18.8 percentage points compared with normal Google Search. The study covered 1,100 US users over seven days in March 2026.

Do AI Overviews reduce traffic too?

The study found that hiding AI Overviews increased click-through to external sites by 8.8 percentage points, which implies AI Overviews reduce clicks. The hiding only worked about half the time, so the authors report adjusted estimates.

Did users prefer AI Mode?

Not on average. Users assigned to AI Mode reported lower trust, usefulness and satisfaction, and 11.2 points more of them searched on Bing, DuckDuckGo or Yahoo. About 29% gave a positive overall impression and 34% a negative one.

Does this study apply to Thailand?

Not directly. All participants were US-based English-language users. The direction of the findings is a reasonable planning input, but the size of any effect in Thai search is unmeasured.

Why did ad clicks fall so much in AI Mode?

Because AI Mode did not show ads during the March 2026 experiment. The authors state this directly, so the 42.7-point drop reflects the absence of ads rather than how people respond to ads in AI Mode.

This is the strongest causal evidence so far that Google's AI search features move clicks away from websites, and that AI Mode does so without winning users over, at least in its first weeks. The results are US-only and short-term. They are still a clear signal to plan for fewer clicks per search and to track AI visibility as its own metric.