AI Search Cuts Publisher Clicks: The Next SEO Move

A preregistered field experiment involving 1,100 Google users found that forcing searches into AI Mode reduced external click-through by 18.8 percentage points, while removing AI search features increased it by 8.8 points. Publishers should now measure AI exposure, human visits, and conversions separately, then build pages that offer value beyond a generated summary.

What did the new AI search traffic study find?

The August 2026 study, AI in Search Reduces Publisher Referrals Without Improving User Experience, tested generative search in everyday browsing rather than asking participants to complete artificial laboratory tasks. Its headline result was substantial: assignment to Google AI Mode reduced click-through to external websites by 18.8 percentage points compared with the current Google Search experience. The 95% confidence interval ran from -22.2 to -15.3 points, and the reported p-value was below 0.001.

In the opposite condition, exposure to a version of Search that attempted to hide AI Overviews and redirect AI Mode increased external click-through by 8.8 percentage points. That estimate needs an important qualification: Google's markup changed during the experiment, so the browser extension's success at hiding AI Overviews deteriorated. The researchers therefore reported a local average treatment effect based on actual exposure, not a simple perfect-compliance comparison.

Assignment to AI Mode also reduced clicks to news, Reddit, and Wikipedia; increased session duration by 0.43 minutes while reducing daily search sessions by 0.92; lowered reported trust and satisfaction; and increased competing-search-engine use by 11.2 percentage points.

These are results from one preregistered experiment, not universal constants. The complete methods and estimates are available in the authors' original arXiv paper.

How was the experiment designed?

Researchers recruited US adults who primarily used Chrome and Google Search. Participants first completed a three-day baseline period, then were randomly assigned to one of three conditions for seven days:

Condition	What participants experienced	What it tests
Current Search	Google's ordinary interface, where AI Overviews could appear and AI Mode remained available	The real-world control condition
No AI Search	AI Overviews were hidden where possible and AI Mode searches redirected to ordinary Search	What changes when generative features are removed
AI Mode Search	Searches were redirected into Google's conversational AI Mode	What changes under intensive AI Mode use

Of 1,444 enrolled participants, 1,100 entered the behavioral analysis and 956 completed the post-experiment survey. Random assignment supports stronger causal inference than a before-and-after traffic comparison, where seasonality, rankings, and audience mix can move simultaneously.

Before treatment, only 0.6% of observed searches used AI Mode; during treatment, 94.7% of searches in that group were redirected there. The estimate describes near-complete seven-day exposure, not voluntary adoption across every market.

What does the evidence prove, and what remains uncertain?

Within its tested setting, the experiment shows that conversational AI search can substitute for external visits. It does not prove that every AI Overview, query type, or publisher experiences the same decline.

Five boundaries matter:

1. AI Mode and AI Overviews were different treatments. The -18.8-point estimate belongs to assignment to AI Mode. It should not be presented as the effect of an individual AI Overview.
2. The sample was not nationally representative. Participants skewed younger, more educated, and more politically left-leaning than the US population.
3. Treatment lasted seven days. People may adapt differently over months, and the product itself continues to change.
4. The no-AI intervention had partial compliance. A Google interface update progressively prevented the extension from hiding AI Overviews reliably.
5. The tested AI Mode did not display ads. Later product changes may alter behavior.

A separate 2026 field experiment by Saharsh Agarwal and Ananya Sen also reported fewer outbound organic clicks when an AI Overview appeared, using a different browser-extension design. The authors report a 39.8% reduction conditional on an Overview appearing, which is a relative percentage, not 39.8 percentage points. The paper is available through its SSRN record.

The studies reinforce the direction of the risk, but their estimates should not be combined: they used different interventions, populations, denominators, and periods.

How does this compare with Google's position?

Google has reported a different aggregate perspective. In August 2025, the company said total organic click volume from Google Search to websites was relatively stable year over year and that the average quality of those clicks had increased. Google defined higher-quality clicks as visits where people did not quickly return to Search and argued that longer, more complex queries can create new opportunities for websites. That claim and its framing appear in Google's article on AI search and higher-quality clicks.

These sources measure different objects. Google's statement describes aggregate traffic and proprietary quality signals; the experiment isolates interface exposure in a randomized sample. Total search volume could grow while click probability falls, and fewer visits could still show stronger engagement.

The useful question is not whether AI search is universally good or bad. It is where a site's exposure, referrals, engagement, and conversions are moving on commercially important pages.

Why AI impressions, clicks, and conversions need separate KPIs

An AI citation is not a visit, and a visit is not automatically a subscriber, lead, or customer. Combining these stages into one score makes losses difficult to diagnose.

Use a four-stage measurement model:

Stage	Useful evidence	Decision it supports
Eligibility	Indexability, crawler access, canonical status, rendered text	Can the page be considered?
Exposure	Google AI impressions, Bing citations, tracked prompt mentions	Is the page or brand appearing?
Engagement	Human sessions, landing pages, scroll or engaged time	Did someone visit and use the page?
Outcome	Sign-ups, leads, subscriptions, trials, purchases	Did the visit create business value?

Google's Generative AI report exposes AI Mode and AI Overview impressions in Search Console, but no dedicated click or prompt dimension. The Search Console AI Performance guide explains its limits. Treat impressions as exposure and use analytics plus successful server-side events downstream.

For other engines, run a fixed prompt set and preserve the answer, citation URL, date, platform, market, and competitors. A free AI visibility check diagnoses readiness; recurring prompts measure observed mentions and citations. The broader GEO framework connects both layers.

How should publishers adapt their SEO and GEO strategy?

Do not hide useful answers until after a click. A vague or inaccessible page may lose both eligibility and trust. Make the extractable answer accurate and the destination substantially more useful than the summary.

Preserve the answer that earns exposure

Open with a self-contained answer, followed by scope and limitations. This gives retrieval systems a faithful passage and lets visitors confirm relevance quickly.

Google's generative AI guidance emphasizes standard SEO eligibility and original, non-commodity content. It does not recommend hiding answers, generating every query variation, or adding AI-only markup.

Create a reason to continue beyond the summary

The click needs a job that the generated answer cannot complete. Depending on the page, that reason might be:

• an interactive checker, calculator, configurator, or diagnostic;
• original data with a transparent methodology and downloadable detail;
• a comparison that exposes criteria, trade-offs, and current pricing;
• a template, spreadsheet, implementation checklist, or code example;
• screenshots or a reproducible test that lets the reader verify a claim;
• expert interpretation for a specific industry, market, or constraint.

Adding 1,000 words that restate the answer only creates more material to summarize. The extra value must be functional or evidentiary.

Optimize the post-click path

When qualified clicks become scarcer, conversion friction matters more. Match the call to action to the intent, place it after the answer, and measure successful actions instead of button presses.

For a diagnostic page, the natural action is a test. Research can offer its methodology; a comparison can let readers evaluate explicit requirements.

Strengthen sources that engines can verify

Publish original findings, method, date, primary evidence, and responsible author. Keep entity information consistent across the site and credible external profiles.

What should a 30-day response plan include?

Start with measurement rather than a sitewide rewrite:

• Days 1-3: Export Google AI impressions, ordinary Search performance, Bing citations, landing sessions, and conversions.
• Days 4-7: Select three exposed, commercially relevant pages and verify indexability, mobile rendering, and event collection.
• Days 8-14: Add one destination-only asset to each page while preserving the direct answer.
• Days 15-21: Improve contextual links and the post-click action. Record the deployment date.
• Days 22-30: Compare complete windows for exposure, visits, engagement, and successful conversions.

Use the AI visibility measurement workflow to keep readiness, observed visibility, and business impact separate. For source-level referrals, the ChatGPT traffic guide provides a practical GA4 and Plausible setup.

What should publishers avoid doing?

Do not replace introductions with sales copy, block crawlers to manufacture clicks, remove useful facts, or mass-produce thin fan-out pages. Those actions reduce discoverability without creating a better destination.

Do not claim causality from one analytics dip. Rankings, seasonality, tracking failures, incomplete data, and interface experiments can move simultaneously. Annotate changes and compare complete periods.

The strategic shift is narrower: clicks from informational discovery may become harder to earn, so each exposed page needs a verifiable answer, distinctive depth, and a measured next step.

FAQ

Did the study find that AI Overviews reduce clicks by 18.8%?

No. The 18.8 figure is a percentage-point reduction in external click-through among participants assigned to AI Mode, relative to Current Search. The separate no-AI condition estimated an 8.8-point increase under partial compliance. Neither number should be relabeled as the universal effect of one AI Overview.

Does lower publisher traffic mean GEO is no longer valuable?

No. Visibility can influence awareness and later branded demand even without an immediate click. GEO remains useful, but exposure, citation, referral, and conversion metrics must be reported separately so visibility is not mistaken for business impact.

Should publishers stop giving direct answers?

No. Hiding the answer can weaken relevance and user experience. Provide a concise, accurate answer, then offer something a generated summary cannot deliver, such as an interactive tool, original evidence, detailed comparison, template, or implementation workflow.

How can a site measure Google AI traffic accurately?

Use Google's Generative AI report for supported-feature impressions, ordinary Search Console data for general clicks, and analytics for human landing sessions and conversions. Google does not currently provide a dedicated AI click or prompt dimension, so do not calculate an official AI CTR from mismatched metrics.

Is this study definitive for every website and market?

No. It is a preregistered randomized field experiment with strong causal value for its tested conditions, but it used a US-based, non-representative sample and a seven-day intervention. Product design, adoption, query mix, geography, and publisher type can produce different effects.

Sources

• AI in Search Reduces Publisher Referrals Without Improving User Experience - Wang et al., preregistered field experiment, August 18, 2026
• The Impact of Google AI Overviews on Publisher Traffic and User Experience - Agarwal and Sen, field experiment, revised July 8, 2026
• AI in Search is driving more queries and higher quality clicks - Google, August 6, 2025
• Optimizing your website for generative AI features on Google Search - Google Search Central
• Introducing Search Generative AI performance reports - Google Search Central, worldwide rollout noted August 31, 2026