Google, Similarweb, Comscore, Semrush and two academic studies reveal reach, visits, substitution, referral and feature activation.
Google says AI Overviews now reaches more than 2.5 billion monthly active users. AI Mode has surpassed one billion. Similarweb estimates that generative-AI websites averaged 9.5 billion visits a month over the 12 months ending May 2026. Comscore counted 76 billion US desktop searches in the first quarter of 2026.[1] [2] [3]
Those numbers make AI search look enormous. They also measure different things.
One counts people encountering AI inside Google Search. One counts visits to standalone AI websites. One counts conventional desktop searches. None tells us what percentage of the information-search market has moved to AI.
That is the updated answer to a question AEO Updates examined one day ago: <strong>AI search is no longer one emerging channel beside traditional search. It is a set of overlapping behaviours inside and outside the search products people already use. Its scale cannot be represented honestly by one market-share percentage.</strong>
<h2>Five measurements, five different questions</h2>
The most useful way to describe the market is to separate five layers.
<div class="overflow-x-auto"><table><thead><tr><th>Measurement layer</th><th>Question it answers</th><th>Current evidence</th><th>What it does not establish</th></tr></thead><tbody><tr><td>Embedded reach</td><td>How many people encounter AI inside conventional search?</td><td>Google reports more than 2.5 billion monthly active users for AI Overviews and more than one billion for AI Mode.[1]</td><td>Unique combined users, query share or a separate AI-search market</td></tr><tr><td>Standalone activity</td><td>How much traffic reaches generative-AI destinations?</td><td>Similarweb estimates 9.5 billion average monthly worldwide website visits during June 2025 to May 2026.[2]</td><td>People, prompts, information searches or all embedded and API use</td></tr><tr><td>Behavioural substitution</td><td>What changes when people gain access to AI search?</td><td>A US desktop quasi-experiment estimated a 9.4% reduction in traditional-search queries after wider ChatGPT Search access.[4]</td><td>ChatGPT market share or a universal cross-device effect</td></tr><tr><td>Referral behaviour</td><td>How often does an AI information-seeking occasion produce a visible handoff?</td><td>The same study found a clean outbound referral in 5.2% of reconstructed ChatGPT conversation sessions.[4]</td><td>Per-link click-through, task completion or unclicked influence</td></tr><tr><td>System incidence</td><td>When does an AI-search feature or live web retrieval activate?</td><td>Studies report rates ranging from 13.7% of one Google query sample to 64.7% of its question queries, while Semrush observed ChatGPT web search on 34.5% of sampled queries in February 2026.[5] [6]</td><td>A permanent system property or universal rate across prompt and query populations</td></tr></tbody></table></div>
These layers overlap. A person can use AI Mode inside Google, visit ChatGPT later, receive a web-search-enabled response, click no citation and then navigate directly to a brand. Counting any one of those events as “AI search market share” hides most of the journey.
<h2>Embedded AI is already operating at search scale</h2>
At Google I/O on May 19, 2026, Sundar Pichai said AI Overviews had more than 2.5 billion monthly active users and AI Mode had surpassed one billion monthly active users in its first year.[1]
Those figures are not additive. A person may use both products, and both sit inside Google Search. They are also first-party product figures rather than independently audited population estimates.
They still matter. Google is not asking billions of people to abandon conventional search before they encounter a generated answer. It is embedding AI into a behaviour they already have.
Google’s April earnings remarks said Search queries were at an all-time high and that AI Overviews were driving overall Search growth.[7] That is Google’s own account of product performance, not causal proof that AI created the growth. But it reinforces an important measurement boundary: growth in embedded AI can coexist with growth in conventional search because they are increasingly part of the same product.
For marketers, this is the largest addressable AI-search surface in the current evidence. It is also the easiest one to misunderstand. An AI Overview user is not necessarily a standalone-assistant user, and an AI Mode user is not a separate addition to Google’s total audience.
<h2>Standalone AI activity is large, growing and still not search volume</h2>
Similarweb’s 2026 Generative AI Landscape describes a different surface. It estimates that generative-AI websites averaged 9.5 billion visits a month worldwide from June 2025 through May 2026, up 70% year over year. Unique visitors rose 57% to 655 million.[2]
The faster growth in visits than visitors suggests that existing users were returning more often. But a visit is not a prompt, and a prompt is not necessarily a search. People use assistants to write, code, analyse, create images, summarise documents and perform many other tasks.
Similarweb also reports that 95% of ChatGPT users used Google.[2] That overlap argues against a simple replacement story. It does not prove that every user moved sequentially from ChatGPT to Google, or that either product caused use of the other. It shows that the audiences are largely shared.
Semrush’s 17-month US clickstream study helps explain why total ChatGPT traffic overstates AI-search activity. It observed ChatGPT’s web-search feature on 34.5% of sampled queries in February 2026, down from 46% in late 2024. Across the study, the rate ranged from 15% to 66.3% as models and product behaviour changed.[6]
That volatility is not a nuisance to average away. It is part of the result. Whether a ChatGPT interaction becomes a live web-search event depends on the prompt, the model, the mode, the date and the product’s retrieval policy.
<h2>There is causal evidence of substitution, but not a market-share census</h2>
The strongest evidence that standalone AI search can displace conventional search comes from a July 2026 preprint by Qiaoni Shi, Kai Zhu and Kai Gu.[4]
The researchers used URL-level Comscore US desktop clickstream from October 2024 through July 2025. They exploited three expansions of ChatGPT Search access and compared 3,882 treated households with reweighted controls in a stacked difference-in-differences design.
Wider access reduced traditional Google, Bing and Yahoo queries by 9.4% relative to the pre-expansion mean among treated households. After 20 weeks, the estimated reduction reached 17%.[4]
This is meaningful causal evidence within the study’s design. It is not evidence that ChatGPT owns 9.4% or 17% of search. The population was a historical US desktop panel. The treatment was access expansion. The outcome was the change in traditional-search queries among affected households.
The distinction between a treatment effect and market share is not technical pedantry. It changes the decision. A marketer can reasonably infer that broader AI-search access may reduce some conventional informational searching. The same marketer cannot use 9.4% as the new share of search belonging to ChatGPT.
<h2>Search is changing without disappearing</h2>
Comscore’s Q1 2026 AI Intelligence Report offers a useful counterweight. It counted 76 billion US desktop searches in the first quarter of 2026, 10% more than in the first quarter of 2024.[3]
The same commercial panel reported that AI assistants reached 36% of US desktop users and 23% of mobile users during the quarter. ChatGPT led with 87 million desktop visitors and 244 million desktop conversations in March.[3]
These figures do not prove that AI expanded the total market. They show that aggregate US desktop search remained large and grew in Comscore’s panel while assistant use also expanded.
That coexistence is easier to understand when the unit changes from a query to a decision journey. Comscore reported an average of 4.9 prompts per ChatGPT conversation, 4.6 on Gemini and 7.1 on Copilot in March 2026.[3] A multi-turn conversation and a keyword query are not interchangeable events. One may replace several searches, create new follow-up questions or send the user back to Google.
<h2>Referral traffic measures the handoff, not the whole influence</h2>
Shi, Zhu and Gu reconstructed 409,133 ChatGPT conversation sessions and 61.5 million Google queries in their US desktop data. A clean outbound referral occurred in 5.2% of ChatGPT conversation sessions, compared with 31.1% of Google queries.[4]
The denominator matters again. The ChatGPT figure is not clicks divided by displayed citations. It asks whether the whole conversation session produced at least one clean outbound visit.
The remaining 94.8% cannot be labelled “successful zero-click answers”. Some sessions may have satisfied the user. Some may have been abandoned. Some later visits may have lost their referrer. The study measures observable routing, not task completion or consumer welfare.
Semrush found that outbound ChatGPT referral traffic grew 206% from January 2025 to January 2026 and that 21.6% of February 2026 referrals went to Google.[6] Those findings again suggest coexistence rather than a clean channel replacement. They also show why referral analytics are a lagging and incomplete view of AI influence.
<h2>AI Overview incidence depends on the query universe</h2>
How often Google displays an AI Overview should be a simple metric. Current estimates show why it is not.
Similarweb reports that more than 40% of US searches triggered an AI Overview in its proprietary measurement.[2] A May 2026 academic preprint by Haofei Xu, Umar Iqbal and Jacob Montgomery found 13.7% activation across 55,393 US trending queries collected over 40 days. Within that same study, 64.7% of question-form queries triggered an AI Overview.[5]
Those figures should not be averaged. The academic sample covered trending queries across 19 categories, not a random census of all search traffic. Similarweb uses its own panel and query-estimation methodology. Query form alone changed activation by nearly five times within the academic sample.
The gap is not evidence that one source must be wrong. It is evidence that the prompt or query universe defines the metric.
This is the same problem that affects AI visibility studies. A brand’s visibility can change when the prompt family, market, model, mode, date or repetition policy changes. The denominator is not a footnote. It is part of the finding.
<h2>A four-denominator operating model for marketers</h2>
The evidence supports four questions for executive reporting. They should be shown together but never collapsed into one composite market-share number.
<div class="overflow-x-auto"><table><thead><tr><th>Executive question</th><th>Appropriate evidence</th><th>Recommended interpretation</th></tr></thead><tbody><tr><td>Adoption: How many people encounter or deliberately use AI for search and discovery?</td><td>Platform-reported embedded reach, unique visitors, representative user surveys</td><td>Addressable audience, with product and overlap boundaries</td></tr><tr><td>Substitution: How much conventional search behaviour changes after AI-search access or use?</td><td>Experiments, quasi-experiments and longitudinal clickstream</td><td>Incremental behavioural change in a named population, not universal market share</td></tr><tr><td>Influence: Where does AI shape research, comparison, recommendation or brand perception without a tracked click?</td><td>Decision-stage studies, matched exposure analysis, surveys and brand lift</td><td>Probable decision influence, clearly separated from observed traffic and conversion</td></tr><tr><td>Referral: How much observable traffic moves from AI interfaces to websites?</td><td>Session-level outbound referrals, analytics and clean-referrer analysis</td><td>Measured handoff, not total influence or task completion</td></tr></tbody></table></div>
Embedded-system incidence should accompany all four. Teams need to know which model, mode, retrieval condition and query or prompt family produced the observation.
<h2>What marketers should do now</h2>
The practical response is not to choose one “correct” AI-search percentage. It is to build a measurement system that prevents unlike percentages from becoming interchangeable.
First, label every number by unit: user, unique visitor, visit, query, conversation, prompt, referral or conversion. If the unit is not visible, the number is not decision-ready.
Second, separate embedded AI from standalone assistants. Google AI Overviews and AI Mode are AI search, but they operate inside the largest conventional-search product. Treating them as a competing external channel misstates the customer journey.
Third, keep substitution separate from reach. A large user base does not reveal how much conventional behaviour changed. A treatment effect does not reveal a platform’s market share.
Fourth, report referral and influence separately. A visible click can be attributed. An unclicked mention or recommendation may still matter, but it requires a different method and a more cautious claim.
Fifth, timestamp system conditions. The share of ChatGPT queries using web search moved materially during Semrush’s study. AI Overview activation also varied sharply by query form. A current rate is a dated observation, not a permanent platform attribute.
Finally, resist the executive request for one all-purpose number. A single percentage is easy to repeat and difficult to defend. A compact four-part scorecard is more honest and more actionable.
<h2>AEO Updates Takeaway</h2>
AI search is already large enough to influence how brands are discovered, compared and chosen. It is not yet measurable as one clean market.
Google has embedded generated answers into a product used by billions. Standalone AI websites attract billions of monthly visits. One quasi-experiment finds real substitution away from conventional search. Referral studies show that many AI conversations do not produce a visible website handoff. Feature-incidence studies show that the result depends heavily on the query universe and system state.
All of those findings can be true at once.
The mistake is not choosing the wrong percentage. The mistake is pretending they are percentages of the same thing.
<h3>References</h3>
[1] <a href="https://blog.google/innovation-and-ai/sundar-pichai-io-2026/" target="_blank" rel="noopener noreferrer">Google, “I/O 2026: Welcome to the agentic Gemini era”</a>, 19 May 2026.
[2] <a href="https://www.similarweb.com/corp/reports/2026-generative-ai-landscape/" target="_blank" rel="noopener noreferrer">Similarweb, “The 2026 Generative AI Landscape Report”</a> and <a href="https://aisearch.similarweb.com/blog/gen-ai-stats/" target="_blank" rel="noopener noreferrer">“AI Search Stats 2026”</a>, July 2026.
[3] <a href="https://www.comscore.com/Insights/Press-Releases/2026/6/Comscore-s-Q1-2026-AI-Intelligence-Report" target="_blank" rel="noopener noreferrer">Comscore, “Q1 2026 AI Intelligence Report”</a>, 2 June 2026.
[4] <a href="https://arxiv.org/abs/2607.07652" target="_blank" rel="noopener noreferrer">Qiaoni Shi, Kai Zhu and Kai Gu, “Answering Without Referring: How AI Search Rewrites the Web’s Economic Bargain”</a>, arXiv preprint, 8 July 2026.
[5] <a href="https://arxiv.org/abs/2605.14021" target="_blank" rel="noopener noreferrer">Haofei Xu, Umar Iqbal and Jacob M. Montgomery, “Measuring Google AI Overviews”</a>, arXiv preprint, 13 May 2026.
[6] <a href="https://www.semrush.com/blog/chatgpt-search-insights/" target="_blank" rel="noopener noreferrer">Semrush, “ChatGPT traffic analysis: Insights from 17 months of clickstream data”</a>, 7 April 2026.
[7] <a href="https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q1-2026/" target="_blank" rel="noopener noreferrer">Google, “Q1 2026 earnings call: Remarks from our CEO”</a>, 29 April 2026.