Most ChatGPT users also visit Google. Similarweb put the share of ChatGPT users who also show up in Google’s audience at 95% in September 2025, and the rate stayed at 95% through May 2026, even as visits to generative AI platforms grew by 70% year-over-year. Similarweb reads that as AI being added to search rather than replacing it. Greg Jarboe highlighted this in his July SEJ review.
A study from Bocconi University, which I covered a month earlier, showed a 9.4% drop in traditional search queries among households that gained access to ChatGPT Search, with the gap growing the longer they had it. Both can be true because overlap counts people, whereas the Bocconi study counts queries per household.
This year’s research keeps showing those two measures moving on their own, along with a third, clicks, which can fall even when the number of people, and queries hold. A search practitioner being asked why referrals fell while reports say nobody left Google is standing in the middle of that gap.
Using Google Is A Low Bar To Clear
Overlap helps us determine if the same individuals are part of two different audiences during a specified time frame. For example, someone who performs a single Google search in a month is categorized as a Google user. This could involve someone who has moved some of their research to ChatGPT but still uses Google for maps, login, and store hours.
Similarweb’s August post on the report explains that overlap looks at whether the same people appear in both audiences, rather than analyzing how they interact in a single research session. This means that someone can still be a Google user while sending a growing share of specific query types to ChatGPT. The report doesn’t specify query-level substitution, and this is confirmed in the post. The report also leaves out API usage, desktop apps, and in-app AI features, which Similarweb says are more likely to replace search activities rather than add to them. For example, a developer using the API or someone doing all their activities through a desktop app wouldn’t be reflected in the panel.
Overlap doesn’t indicate how frequently people use these services. Pew’s February survey of 5,119 U.S. adults revealed that 49% have ever used a chatbot, 42% use one to find information, and 60% read AI summaries at the top of search results. I covered these statistics in June. “Ever use” is a broad metric, and these figures don’t detail how usage is distributed across various activities.
Same Users, Fewer Searches
This Bocconi paper explores how search behavior changes when households gain access to ChatGPT Search. Using Comscore’s U.S. desktop clickstream data from October 2024 to July 2025, a window covering OpenAI’s three expansions of ChatGPT Search access, the study compares households that gained access with similar households that had no prior ChatGPT or Claude activity.
Results show that, on average, households reduced their traditional search queries by 3.14 per week after gaining access, which is roughly 9.4% less than the pre-expansion average of 33.51. This decline increased over time, reaching 17.0% after 20 weeks.
Comparing logged-in ChatGPT users with Search access in December 2024 to anonymous users who gained access in February 2025 revealed smaller declines, 4.9% on average, and 8.2% after 20 weeks. The authors read the smaller figure as a sign that comparing adopters with nonusers may pick up differences in who adopts AI rather than the effect of access.
The biggest drops in search referrals were in informational categories. Referrals to academic sites dropped by 32.8%, reference sites by 26.5%, developer sites by 15.1%, and news sites by 13.4%. In contrast, referrals to marketplaces, and entertainment sites barely moved or couldn’t be told apart from zero. The paper emphasizes that these patterns are descriptive and do not establish causality.
According to the same Comscore panel, Google’s household reach remained steady at about 49%, while ChatGPT’s expanded from 4.6% to 7.9%. Households with access still ran fewer queries, so Google’s audience held while queries fell, in one dataset. These insights come from a working paper analyzing desktop browsing data up to July 2025, before AI Mode reached its current scale.
Other Research Finds Search And AI Used Together
A working paper by Samira Gholami, Cristiana Firullo, Cristobal Cheyre, and Alessandro Acquisti on SSRN in February reports that users who used a chatbot visited more websites afterward. They studied 4,394 Comscore desktop users who started using a large language model (LLM) between August 2023 and January 2024. Adoption was linked to a consistent increase in weekly unique website visits and a temporary boost in traditional search activity, which later declined.
The authors call the search estimates noisy and stop short of claiming a lasting reduction. Data from a 2025 field study with 308 U.S. Chrome users points the same way, with sessions that used both an LLM, and search engine lasted longer, and involved more sites, while LLM-only sessions were shorter, and narrower. Since the Comscore data predates ChatGPT Search and AI Mode, it pertains to earlier products than those discussed by Bocconi.
The team behind the AI Mode experiment reviews these earlier studies in their own paper and describes the evidence as mixed, with some studies showing people search less after adopting a chatbot while others indicate they use both side-by-side.
AI Mode Can Replace The Results Page Without Replacing Google
Google’s numbers make it harder to treat Google and AI as two separate things to compare. At I/O in May, Sundar Pichai announced that AI Overviews had over 2.5 billion monthly active users, while AI Mode hit over 1 billion in its first year. Liz Reid also noted record query volumes last quarter. Pichai summarized the company’s perspective: “When people use our AI-powered features in Search, they use Search more.”
These claims don’t have baseline or user-level data for independent verification. What the numbers do show is that a growing share of Google use runs through AI features, which I covered when the first AI Mode usage data came out. When users engage with AI Mode, they are both Google, and AI users, and may never see the traditional results page.
A recent field experiment, published on arXiv on August 18, looked into what happens when AI Mode becomes the main experience. Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson, and Danaé Metaxa recruited 1,100 U.S. Chrome users who use Google as their main search engine. After observing their activity for three days, they randomly assigned participants to three groups for a week: standard Google, Google with AI Overviews hidden, or Google with all searches conducted in AI Mode.
At baseline, 0.6% of searches used AI Mode, but in the treatment group, 94.7% of searches were routed there, which the authors describe as forced, full adoption rather than the gradual opt-in most people would experience. The researchers also noted that the effects of voluntary use may differ.
In this setup, there was an 18.8 percentage point decrease in click-through to external sites (with a 95% confidence interval of -22.2 to -15.3), and search sessions declined by roughly 0.92 per day, from an average of 4.0. Notably, more frequent Google users experienced a larger decline. Meanwhile, the average session duration increased by 0.43 minutes. The share of participants clicking through to news sites dropped by 12.5 points, Reddit by 21.2 points, and Wikipedia by 9.9 points.
Participants also looked at other options. The share of participants using Bing, DuckDuckGo, or Yahoo went up by 11.2 points, but ratings for trust, satisfaction, and usefulness went down. Out of 309 AI Mode users who answered an open-ended question, 47 (about 15.3%) reported difficulty reaching specific websites, and one person mentioned that when they entered a site, “it would tell me about the website instead.” When AI Overviews were hidden, external clicks increased by 8.8 points among those where the extension worked. However, a mid-study change to Google’s interface disrupted the hiding feature for roughly half of the AI Overviews.
The study primarily involved younger, more educated, and more left-leaning users compared to the general U.S. population. Additionally, a seven-day period is insufficient to observe long-term habit changes. In April, I discussed an experiment by Agarwal and Sen that also routed a group into AI Mode and found fewer outbound clicks and lower satisfaction, though those authors labeled that arm exploratory because some participants dropped out. Wang’s team pre-registered their AI Mode comparison and found no evidence of increased dropout, so the two experiments now point the same way.
Why This Matters
A news website can maintain consistent Google impressions despite a decline in referral traffic, aligning with research findings. Bocconi reported a 13.4% decrease in search referrals to news sites among households that gained access to ChatGPT Search, while Wang observed a 12.5-point reduction in users clicking on news links under AI Mode assignment. In Bocconi’s study, Google’s audience remained stable even with fewer searches, whereas Wang’s data indicated users continued searching on Google but clicked on fewer links. The pattern was different for marketplace and entertainment sites in Bocconi’s results. Referrals to these sites stayed almost the same after ChatGPT Search access, but referrals to reference, and academic sites dropped by more than 25%, and the paper stops at describing that split. Wang’s study also found some people had trouble reaching specific websites of any kind under AI Mode. When looking at all traffic together, these differences aren’t as clear.
Agencies frequently hear the 95% figure from clients, usually based on headlines. Similarweb views the continued overlap as an indication that people haven’t left Google altogether, at least among the audience it monitors. However, this number doesn’t reveal how many queries those users are making. The closest thing to a controlled test of that found traditional search queries declined on U.S. desktops after ChatGPT Search access, particularly in informational categories.
When someone asks ChatGPT for a recommendation and then searches for that brand on Google, it appears as an organic visit in Search Console. However, there’s no straightforward way to determine the actual source of that visit. That is the in-house team’s problem, and the overlap number hides it. Similarweb names branded search lift as one of the metrics that would move if AI were feeding demand, and calls it a type of conversion that most analytics tools don’t typically attribute to an AI session.
Overall, metrics like audience size, query volume, search sessions, AI visibility, and referral traffic are often discussed as if they’re part of the same measurement. This year’s data shows that, while Google keeps users engaged, it’s losing some queries, and it can keep the query, and still lose the click. In Wang’s study, people kept searching while clicks, and trust fell.
What None Of These Datasets Can See
None of the datasets we’ve discussed before track individual users across various platforms like Google Search, AI Mode, ChatGPT, and Gemini over a long period at the task level. To give some background, Similarweb mainly concentrates on audience data rather than detailed session information. Bocconi and Gholami analyze U.S. desktop clickstream data, with Bocconi’s dataset ending in July 2025. Wang’s experiment follows around a thousand Chrome users over seven days, during which their searches were routed into AI Mode.
Currently, Google has not released user-level data to verify its claims. Its I/O posts highlight record-high query volumes and increased search activity but do not specify what constitutes a query in AI Mode, provide a baseline, or clarify which queries have increased.
Additionally, it’s unclear which types of searches are moving to AI, and which stay in a search box. Bocconi’s split between informational and transactional categories is descriptive only. Wang’s team reminds us that their experiment wasn’t set up to pick up small effects in secondary click outcomes, and Gholami’s session categories are derived from data gathered from 308 users.
Looking Ahead
During Wang’s March experiment, AI Mode didn’t show ads, but the authors expect satisfaction, and click metrics to move once ads are introduced. On May 20, Google shared that they are testing two new ad formats created with Gemini in AI Mode and expanding their Direct Offers pilot to include responses generated by AI Mode. The earlier data on clicks and trust come from a previous product version. Remember, all three academic papers are still drafts, and haven’t yet been peer-reviewed. I’ll update if any of the findings change in review. Alphabet’s upcoming earnings call, usually in late October, is where Google is likely to update its query-growth claim, and the thing to listen for is whether it comes with a baseline this time.
More Resources:
- Google CEO Sundar Pichai Is OK With AI Mode Replacing Classic Search
- Google Data Compares Gemini & AI Mode Use Against Daily Life
- 846,000 Google Searches Reveal How AI Overviews Are Changing User Behavior
Featured Image: Tada Images/Shutterstock
