Do Google AI Overviews Reduce Website Clicks? What the Experiment Actually Proves

Separate the two search experiences first

The August 2026 arXiv preprintstudies three conditions: attempted removal of AI features, unchanged Google Search and forced AI Mode. Its behavioral sample included 1,100 US users. Treatment lasted seven days after a shared three-day baseline. It is a preprint and has not been peer reviewed.

The design graphic distinguishes those assignments. Removing features for a user is different from excluding one publisher's content while the features remain available.

This explanatory graphic describes the research design. It is an illustration, not a Google screenshot or a new experiment.

Read the AI Overview estimate with its uncertainty

The table keeps the original units and separates the question answered by each estimate. Positive values mean more external clicking under the No AI condition; the negative AI Mode value means less clicking.

The chart's No AI LATE row represents a continuous exposure model, not a simple comparison of everyone whose blocker worked against everyone else's.

Intent to treat, or ITT, keeps people in their assigned groups. It answers what happened when the treatment was assigned, including imperfect delivery. An interval that crosses zero does not establish either a positive effect or no effect. It signals uncertainty at the stated confidence level.

Ask what a publisher can actually verify

The useful next step is to separate a research estimate from your site's exposure, actual visits and qualified outcomes. A page appearing in an answer establishes visibility; it does not show that the user clicked, or would have clicked without the answer.

Xtrusio can record your agreed buyer questions, the answers Google returns and the URLs it cites, so you can line them up with later changes in your own data. Monitoring cannot recreate the study's random assignment or supply private site analytics that have not been provided.

Keep those records alongside your own search and analytics data. They can help explain which page or claim to inspect, while leaving the traffic effect open. This prevents a visible citation from being counted as a recovered visit or a completed article from being counted as a business result.

Why is the compliance-adjusted result conditional?

Local average treatment effect, or LATE, asks about exposure changed by assignment under additional assumptions. In this paper, the blocker hid 51.1% of AI Overview exposures overall. The authors used continuous compliance and excluded users with no AI Overview exposures from that LATE analysis outside preregistration. The model assumes assignment affects the outcome through the exposure being estimated; the paper cautions that a broader interface change could violate that assumption.

Check what goes into the denominator

The study's click-through measure divides each user's external Google clicks by Google searches, excluding vertical searches and Google destinations. Its main denominator counts AI Mode turns separately. A conversation-based check still found a reduction, but at 14.9 percentage points. The methods and robustness tables specify these choices.

A Search Console ratio instead divides reported site clicks by reported site impressions. An analytics conversion rate might divide qualified actions by sessions. All can be useful, but changing the denominator changes the question. A decline in one rate cannot be pasted onto another measure simply because both are called a percentage.

For a purely hypothetical illustration, a rate falling from 40% to 31.2% is down 8.8 percentage points, or 22% relative to its starting value. That arithmetic is not the experiment's observed publisher baseline or a forecast. Multiplying a site's traffic by 8.8% would answer neither question correctly.

A publisher worksheet can change the decision

The following numbers are invented for teaching. Assume the same landing-page cohort, country/device scope, complete comparison periods and consistently measured qualified action. No site test, customer result or AI-caused change is claimed.

More clicks would not settle this publisher's concern. The next checks concern intent mix, tracking consistency, landing-page experience and action quality. The worksheet does not identify AI as the cause. A demand change, ranking shift or different visitors could produce the same pattern.

Define what counts as a qualified action before comparing periods. A form submission, valid sales inquiry and subscription are different outcomes. If analytics cannot reliably join the outcome to these organic visits, leave the business layer unknown rather than multiplying an estimated click effect by an assumed visitor value.

Use Google's report for the layer it measures

Google's current generative AI report documentationdescribes impressions from AI Overviews and AI Mode together. Pages, countries, devices and dates supply scope; the report does not provide an AI Overview-only click rate. Google dates the report's worldwide rollout to August 31, 2026.

The screenshot below is from Xtrusio's own Search Console property on September 28. It shows the selected three-month, Web (text) report with 248 impressions and Pages, Countries, Devices and Days. The chart covers August 2 to September 25. The account avatar is blurred; report values are unchanged.

This is a dated property-level observation. It contains no exact buyer question, split by AI Overview versus AI Mode, external-click count or conversion result. It therefore illustrates the exposure layer without proving click loss or the effect of a content edit.

Keep aggregation consistent. Google distinguishes property totals from page-level table counts, and report dates use Pacific Time. Its exports can convert unavailable display markers into zeros; preserve their status before treating a zero as measured absence. These reporting details matter more than selecting a convenient multiplier.

The original graphic separates exposure, click yield and business value. It is a conceptual model, not evidence that multiplying three terms identifies an observed causal business impact.

Opting out tests a different intervention

Google's current control documentation says exclusion prevents the site's links and content appearing in covered generative features. Other sources remain available. The control does not act as a ranking or inclusion signal for other Search results, and it is separate from AI training controls.

That is different from hiding the feature for users. Excluding one publisher does not remove the answer or promise that the user will visit that publisher through classic results instead. No exclusion was performed or observed for this article. The experiment alone cannot justify an opt-out recommendation for every site.

Google's August 2025 traffic statement describes aggregate organic clicks as relatively stable. Its quality-click definition concerns users not quickly returning to Search, not a verified lead or purchase. That platform observation and a randomized interface comparison answer different questions; neither establishes your site's outcome.

Decide which page deserves investigation

Choose a cohort whose reader intent and business purpose are clear. Use complete comparable periods and retain a change log for rankings, content, templates, tracking and control settings. A matched reference cohort can help diagnosis, but it does not automatically randomize exposure or remove confounding.

For an answer-complete page, inspect whether the useful next step survives the summary. For a research or transaction page, check whether the visit delivers current evidence, a tool or a decision the reader still needs. Improve that specific value; do not promise that adding it recovers an experimental percentage of traffic.

The AI Overview tracking guide covers question and source inspection; the GEO measurement guide supplies a wider metric dictionary. For your own agreed questions and available evidence, request a scoped free Xtrusio evaluation.