AI Didn’t Replace Google. Your Clicks Still Dropped.

TL;DR: Google’s audience overlap with ChatGPT sits at 95%, but that number masks a real problem: households with ChatGPT Search access ran 9.4% fewer traditional queries within weeks, and a forced AI Mode experiment cut click-throughs to external sites by 18.8 percentage points. Users staying on Google does not mean they are clicking to your site.

The Overlap Statistic Is Technically Correct and Operationally Useless

Similarweb tracked 95% overlap between ChatGPT users and Google’s audience from September 2025 through May 2026. Generative AI platform visits grew 70% year-over-year during the same window. Media coverage ran with the overlap figure as proof that AI isn’t threatening Google. That framing is misleading for any operator who measures performance by referral traffic, lead volume, or cost-per-acquisition.

Overlap counts people. It answers: “Is this person in both audiences this month?” A user who opens Google once to check store hours counts as a Google user, even if they moved every informational query to ChatGPT. Overlap says nothing about query frequency, session depth, or whether that person ever clicked through to an external site. If you run paid media campaigns that depend on organic assist traffic from Google, a 95% overlap tells you almost nothing about why referrals fell.

Similarweb’s own report notes that API usage, desktop apps, and in-app AI features are excluded from the panel. Those channels are, by their own assessment, more likely to substitute search activity than add to it. The overlap number is built on a panel that systematically excludes the behavior most likely to cannibalize your traffic.

Queries Per Household Are Falling, and the Drop Accelerates

The Bocconi University study is the more useful dataset for operators. Researchers used Comscore U.S. desktop clickstream data from October 2024 through July 2025, comparing households that gained ChatGPT Search access against similar households with no prior ChatGPT or Claude activity.

The average household ran 3.14 fewer search queries per week after gaining access, roughly 9.4% below the pre-expansion baseline of 33.51. After 20 weeks, the decline reached 17.0%. The substitution is not a one-time adjustment. It deepens the longer households have the tool.

The category breakdown is what should concern operators in information-heavy verticals. Search referrals to academic sites dropped 32.8%, reference sites fell 26.5%, developer sites declined 15.1%, and news sites dropped 13.4%. Marketplaces and entertainment sites were essentially flat. If your funnel depends on users researching a topic before converting, you are in the categories taking the hardest hits.

Google’s household reach held steady at about 49% in the same Comscore panel, while ChatGPT’s expanded from 4.6% to 7.9%. Audience stable, queries falling. The distinction matters enormously when you are optimizing audience targeting or trying to explain a traffic decline to a client.

AI Mode Is Google. That Is the Problem.

The framing of “Google vs. AI” becomes less useful once you account for AI Mode. At Google I/O in May 2026, Sundar Pichai reported that AI Overviews had over 2.5 billion monthly active users and AI Mode crossed 1 billion in its first year. Liz Reid noted record query volumes in the same quarter.

A controlled field experiment published on arXiv in August 2026 tested what happens when AI Mode becomes the default. Researchers recruited 1,100 U.S. Chrome users and randomly assigned them to standard Google, Google with AI Overviews hidden, or full AI Mode for one week. In the AI Mode group, 94.7% of searches were routed through it. The result: an 18.8 percentage point drop in click-throughs to external sites, and search sessions declined by roughly 0.92 per day from a baseline of 4.0.

Click-throughs to news sites dropped 12.5 points. Reddit referrals fell 21.2 points. Wikipedia lost 9.9 points. About 15% of AI Mode users reported difficulty reaching specific websites. One participant noted that searching for a site returned a summary about the site instead of the site itself.

A separate experiment by Agarwal and Sen found similar results: fewer outbound clicks and lower satisfaction under AI Mode. The two studies now point the same direction. When AI Overviews were hidden in the arXiv experiment, external clicks recovered by 8.8 points. That gap between AI Mode on and AI Mode off is the cost of Google’s current default interface.

What This Means for High-CAC Verticals

Operators in forex acquisition, iGaming player acquisition, and law firm lead generation spend more per acquired customer than almost any other digital advertiser. Organic search assist traffic has historically been a cost offset: a user finds you via informational content, re-engages via paid, converts. That assist layer is shrinking in the categories where AI answers are most complete.

Informational queries, “how does CFD margin work,” “which online casino accepts US players,” “do I have a case for a mass tort” are exactly the query types that AI handles well and users are substituting most aggressively. The Bocconi data showed the largest referral drops in informational and reference categories. Those are the top-of-funnel content types that high-CAC operators have built entire SEO programs around.

The practical implication is not to abandon organic. Google’s audience is still massive and query volume is still high in aggregate. The implication is that impression counts and keyword rankings are no longer reliable proxies for traffic or lead volume. A page can hold its rankings and lose referrals because the AI Overview above it resolves the query without a click. If you have not done a full-funnel marketing audit that separates impressions from click-throughs from assisted conversions, you are making budget decisions on incomplete data.

Operators also need to be honest about where they have structural advantages. Transactional queries, account opens, deposit pages, form submissions, are largely unaffected in the Bocconi data. Marketplaces and entertainment referrals barely moved. Build content architecture that pushes toward transactional pages faster rather than holding users in informational content that is increasingly answered by AI before a click happens.

Three Numbers Worth Tracking Instead of Overlap

The overlap statistic is not going away. It will keep being cited as evidence that Google is fine. Here are three measures that tell you more about what is actually happening to your traffic.

First, track click-through rate by query type, not just aggregate CTR. Informational queries are losing clicks faster than navigational or transactional ones. If your keyword mix is heavy on informational, your CTR decline will outpace your impression decline. Second, track referral traffic by content category. The Bocconi and Wang studies both show category-level variance. News and reference content is under more pressure than marketplace content. Know which bucket your pages sit in. Third, track session quality from organic rather than just session volume. The arXiv experiment found AI Mode increased average session duration on Google by 0.43 minutes. Users are not leaving, they are staying on Google longer and clicking out less. If your organic sessions are shorter and bounce rates are rising, that is consistent with receiving lower-intent clicks from a filtered-down pool.

For operators already using AI-assisted lead qualification on the conversion side, the implication is that the leads arriving from organic may be further along in their research cycle because AI resolved their early questions. Qualification logic should account for that shift. Leads who click through in an AI Mode environment have already been pre-filtered by the AI summary. They arrived with a specific intent. Treat them accordingly.

The crypto operator market is a useful parallel here. Crypto audiences are heavy desktop users who research aggressively before converting, exactly the user profile Bocconi’s data describes as reducing queries most. If your organic strategy for any information-heavy vertical has not been stress-tested against the query substitution data from this research, it needs to be.

Originally reported by Search Engine Journal, September 2026.