The headline from the latest AI search data sounds reassuring for Google: 95% of ChatGPT users still show up in Google’s audience, a figure that held steady from September 2025 through May 2026, according to Similarweb’s August report. But audience overlap and query volume are two different measurements. Right now, they are moving in different directions, and the gap between them is where referral traffic disappears.
What the AI Search Audience Overlap Figure Actually Measures
Overlap, as Similarweb explains in their August post, tells you whether the same individuals appear in two platforms’ user pools within a given time frame. It does not capture how often people use each service, which tasks they assign to each, or whether their behavior inside Google has changed. Someone who now routes most research queries to ChatGPT but still opens Google for maps and store hours counts as a Google user. That is why the 95% figure can hold flat while other numbers shift.
Pew Research’s February survey of 5,119 U.S. adults adds texture: 49% have ever used a chatbot, 42% use one to find information, and 60% read AI summaries at the top of search results. “Ever used” is a broad threshold that does not distinguish daily habits from occasional experiments.
Similarweb also notes a data gap in their report: the panel excludes API usage, desktop apps, and in-app AI features. Similarweb states these channels are more likely to replace search activities than add to them. A developer querying AI through an API, or someone working entirely within a desktop AI app, would not appear in the overlap count at all.
Where the Query Data Tells a Different Story
A Bocconi University working paper, using Comscore U.S. desktop clickstream data from October 2024 to July 2025, compared households that gained ChatGPT Search access with similar households that had no prior ChatGPT or Claude activity. On average, households with access ran 3.14 fewer searches per week, a 9.4% reduction from their pre-access average of 33.51 weekly queries. The gap widened over time, reaching 17.0% after 20 weeks.
Referral traffic by content category showed the sharpest drops in informational content:
- Academic sites: referrals down 32.8%
- Reference sites: down 26.5%
- Developer sites: down 15.1%
- News sites: down 13.4%
- Marketplace and entertainment sites: minimal change, within margin of error
In the same Comscore panel, Google’s household reach held at roughly 49% while ChatGPT’s expanded from 4.6% to 7.9%. Audience numbers stayed stable for Google while query volumes fell, two measurements from a single dataset pointing in different directions. The Bocconi authors note these patterns are descriptive and do not establish causality.
The study also compared logged-in ChatGPT users with search access in December 2024 against anonymous users who gained access in February 2025. The logged-in group showed smaller average declines: 4.9% on average and 8.2% after 20 weeks. The authors read this as a sign that comparing early adopters with non-users may partly reflect who chooses to adopt AI, rather than the pure effect of having access.
Research That Points the Other Way
Not all evidence points toward substitution. A working paper posted on SSRN in February, by Gholami, Firullo, Cheyre, and Acquisti, tracked 4,394 Comscore desktop users who adopted a large language model (LLM, meaning a text-based AI system like ChatGPT or Claude) between August 2023 and January 2024. Adoption correlated with more weekly unique website visits and a temporary rise in traditional search activity that later leveled off. The authors flag the search estimates as noisy and stop short of claiming a lasting reduction.
A 2025 field study with 308 U.S. Chrome users found that sessions combining an LLM and a search engine lasted longer and involved more sites than LLM-only sessions, which were shorter and narrower. Because this Comscore data predates ChatGPT Search and AI Mode, it reflects earlier product behavior than the Bocconi analysis covers.
The team behind Google’s AI Mode research reviewed these earlier studies and described the overall evidence as mixed.
Google’s Own Numbers Blur the Comparison
At Google I/O 2026, CEO Sundar Pichai announced that AI Overviews had surpassed 2.5 billion monthly active users, and that AI Mode, described as “our biggest upgrade to Search ever,” crossed 1 billion monthly active users in its first year. Liz Reid also noted record query volumes last quarter. Pichai’s framing: “When people use our AI-powered features in Search, they use Search more.”
These figures come without baseline comparisons or user-level data available for independent verification. But they point to a structural shift that makes the Google-versus-AI framing harder to apply: for a large and growing share of searches, AI-generated answers are now delivered inside Google, not instead of it.
What This Means for AI-Search Visibility
The practical insight for anyone managing a website or content strategy: audience reach and referral traffic are now two separate dials, and a brand can hold one steady while the other falls. Most traffic reports fail to explain what is happening in that gap.
Consider a how-to blog or a news site. It can still appear in Google’s audience numbers because some readers open Google once a month for something unrelated. Meanwhile, the informational queries that used to drive its traffic are now being answered by an AI overview or a chatbot, with no click required. The audience overlap figure looks fine. The traffic report does not.
The Bocconi referral data makes the exposure clear. A 32.8% drop in clicks to academic sites and a 26.5% drop to reference sites, within the study window, is a large signal for publishers and brands whose content strategy depends on “what is,” “how to,” or “explain this” queries. Those are exactly the query types AI systems answer most confidently and where a generated summary is most likely to satisfy the user without a click-through.
Hingewise’s read: the Similarweb overlap figure is not wrong, but it measures the wrong thing for content strategy decisions. The more useful question is no longer “are users still in Google’s audience?” but “is our content being cited in AI-generated answers?” These are increasingly separate goals. A brand that optimizes only for traditional click-through may maintain its audience overlap number while losing what some practitioners track as Share of Model, meaning how often an AI system cites or recommends your content when answering relevant queries. That is a gap the 95% overlap figure cannot detect. The evidence remains genuinely mixed, and the Bocconi authors are right to avoid causal claims. What is clear is that audience reach, query volume, and referral clicks have decoupled enough that tracking only one gives an incomplete picture. For content teams, the pressing question is which of those three numbers the current strategy is actually built to protect.
Before Drawing Conclusions from Your Traffic Data
- Check whether your analytics separates AI Overview clicks from standard organic results
- Identify whether referral drops are concentrated in informational vs. transactional content categories
- Compare audience reach (unique visitors) against click-through rates to see if they are diverging
- Test manually whether your content appears in AI Overviews or AI Mode responses for key queries
- Confirm whether your traffic measurement is desktop-only (like Comscore) or includes mobile and in-app traffic
- Check whether query declines in your vertical match the informational-heavy pattern in the Bocconi findings
- Establish a pre-AI Mode baseline so you can separate platform effects from your own content performance
The data point worth watching: whether referral declines in informational categories continue to deepen as AI Mode scales further in the U.S., and whether Google eventually releases query-volume data with enough granularity for researchers to verify its record-volume claims independently.
Sources: Search Engine Journal (Matt Southern, July 2026); Google I/O 2026 keynote (Sundar Pichai, May 2026); Similarweb audience overlap report (August); Bocconi University working paper, Comscore U.S. desktop data through July 2025; SSRN working paper, Gholami et al. (February); Pew Research survey (February).
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.
