How Is AI Changing Americans’ Online Experience?

Key Takeaways

• AI references reached nearly every participant, yet deep engagement remained uncommon.
• Search summaries exposed users to generated answers before they actively sought AI tools.
• Publishers now compete with answer pages that can satisfy intent without an outbound click.

What Browsing Data Reveals About AI Online

In March 2025, 900 U.S. adults generated 2,457,176 tracked webpage visits across personal computers, tablets, and phones. The Pew Research Center study used those records to examine how people encountered artificial intelligence (AI) online, including brief mentions, search-result summaries, product descriptions, news coverage, and visits to generative AI tools. The study offers an unusually direct view of behavior because it measures visited pages rather than relying only on survey recall.

The broadest measure found that 93% of participants visited at least one page containing an AI-related term. Yet only 7% of all recorded page visits went to a page with such a term, and about 6% of distinct pages contained one. AI had broad reach without dominating the month’s browsing. A person could encounter the words “AI,” “ChatGPT,” or “AI-powered” in a sidebar, a product listing, a market story, or a search module without choosing to learn about artificial intelligence.

Pew separated incidental exposure from substantive exposure. A substantive page placed AI near the center of the content or indicated active interest, such as an AI-related search, an AI chatbot interface, or an article devoted to the technology. Forty-nine percent of participants visited at least one such page, but substantive visits accounted for less than 0.05% of all page visits. Among people who reached any substantive page, the median was three visits during the month. That gap between population reach and individual frequency is the study’s defining result.

The main measures can be read together rather than as competing descriptions of usage.

The findings describe exposure, not comprehension, approval, dependence, or accuracy. A page-level keyword match does not prove that a participant noticed the term. A visit to a chatbot website does not reveal what the person asked or whether the answer was useful. The study is strongest when used to map where AI entered ordinary browsing and how often users reached pages that treated it as a central subject.

AI Becomes Background Language More Often Than a Chosen Destination

The median participant who encountered any AI-related page did so 60 times during March 2025. That figure sounds high until it is compared with the median of three substantive visits among those who reached substantive content. The larger number captures repeated exposure to labels, modules, sidebars, descriptions, headlines, and search features. The smaller number captures intentional use or sustained attention.

Generic wording carried most of that exposure. Forty-four percent of participants visited a page containing “AI,” 27% saw language such as “AI-powered,” “AI-assisted,” or “AI-enhanced,” and 22% reached a page mentioning ChatGPT. Terms such as “artificial intelligence,” “OpenAI,” “chatbot,” “GPT,” and “generative AI” also appeared often enough to reach meaningful shares of the sample. The chart on page 8 of the Pew study shows that broad descriptors and marketing phrases reached more people than specialist terms associated with model architecture, training, or retrieval.

This language pattern matters because it shows how AI became a commercial adjective. “AI-powered” can describe a product feature, an automated recommendation, a review summary, a customer-service assistant, or a marketing claim with limited technical detail. A consumer may see the phrase many times without learning what model is used, what data it processes, how much human review occurs, or how performance was tested.

The same pattern complicates public-opinion research. Survey respondents may report familiarity with AI after encountering the term in daily browsing, yet their mental models can differ sharply. One person may think of a chatbot. Another may think of a camera feature, a search summary, fraud detection, or a recommendation engine. Exposure to the label creates recognition, but recognition does not establish technical understanding.

Pew found no large differences by age, education, or gender in whether participants saw pages containing AI terms. That result does not mean every group used the same tools or interpreted them in the same way. It indicates that incidental exposure was broad enough to cross demographic lines within the sample. The more selective behavior appeared in direct tool use: 20% of adults ages 18 to 29 visited an AI chatbot website, compared with 13% of all participants.

A useful distinction follows from the data. AI adoption measures deliberate use, such as opening a chatbot or applying an AI feature. AI exposure includes passive contact with generated summaries, automated product text, interface prompts, and brand language. Public discussion often treats those categories as interchangeable, which can overstate active use or understate how deeply AI has entered routine web design.

Search Engines Place Generated Answers Before Deliberate AI Use

Search was the clearest route through which people encountered AI online without asking for it. Fifty-eight percent of participants reached at least one Google results page containing an AI-generated summary. Sixty-five percent saw an AI reference somewhere on a search results page, but only 10% entered a query directly related to AI. The interface introduced AI far more often than users sought AI as a topic.

That distinction changes the meaning of “AI user.” A person can receive a generated answer through ordinary search, read it, and act on it without visiting a chatbot website or identifying the experience as chatbot use. Search systems place model output inside an established habit, reducing the need for a separate adoption decision.

Pew’s click analysis examined what happened after Google displayed an AI summary in the same March 2025 browsing records. Users clicked a traditional search result in 8% of visits with a summary, compared with 15% of visits without one. Links inside the generated summary received a click in 1% of visits. A browsing session ended after 26% of summary pages, compared with 16% of pages containing traditional results only.

Those figures show associations rather than proving that the summary caused every difference. Query complexity, subject matter, and user intent can also affect whether someone clicks a result or ends a session. The pattern still indicates that answer-rich search pages corresponded with fewer outbound visits during the measured period.

A Pew survey conducted in February 2026 found that 60% of U.S. adults said they read AI summaries at the top of search results. Thirty percent said they did not, and 10% were unsure. That survey cannot be compared directly with the 2025 tracking study because one measured self-reported reading and the other measured exposure to visited pages. Taken together, they show that generated search summaries had moved from a limited interface feature into a familiar information format.

Google reported during Google I/O on May 19, 2026 that AI Overviews had more than 2.5 billion monthly active users and AI Mode had exceeded 1 billion monthly active users. Those are company-reported global usage measures, not independent estimates of satisfaction, accuracy, or repeat engagement. They still establish the scale at which Google Search now combines generated text, source links, follow-up questions, commercial features, and advertising.

The commercial consequence is often described as Google Zero, a shorthand for a search environment in which the results page satisfies more intent before a person reaches an outside site. The term does not mean outbound links disappear. It describes a bargaining shift: source pages supply material, search systems synthesize it, and fewer users may need to leave the platform.

News, Shopping, and Social Platforms Turn AI Into Ambient Infrastructure

News pages brought AI terms to 52% of participants, yet only 8% visited a news page where AI received meaningful discussion. A stock-market story might mention Nvidia graphics processors. A policy story might refer to copyright and synthetic media. A sidebar could display an unrelated headline containing “ChatGPT.” These placements count as exposure even when the main article concerns another subject.

That difference matters for news literacy. Seeing an AI-related word near journalism is not the same as reading reporting about model design, labor effects, copyright, safety, competition, or public policy. The study’s page-level approach captures the entire visible text that researchers could retrieve, including recommendation modules. A participant may have focused on the main story and ignored a sidebar. Exposure estimates should not be read as proof that half of adults consumed AI journalism in depth.

Shopping sites produced a similar result. Fifty-four percent of participants visited at least one page on 18 large retail domains that contained an AI reference. Product descriptions promoted AI features, and some pages displayed generated summaries of customer reviews. Amazon accounted for many visits because 63% of participants reached the domain during the month and the retailer had review-summary features on many product pages.

Generated review summaries change the purchase path in a subtle way. They compress large collections of customer comments into a small set of recurring themes, which can save time. They also place trust in the system that selects, weighs, and phrases those themes. A summary can omit disagreement, reduce minority experiences, or give a polished impression that exceeds the underlying evidence. Consumers need accessible routes to the original reviews and clear labeling that distinguishes generated text from customer-authored text.

Social media created the broadest category-level reach in the study: 75% of participants visited a social page containing an AI reference. Much of that result came from Facebook pages that displayed a Meta AI link in a sidebar. It should not be interpreted as evidence that three-quarters of participants actively used Meta AI or consumed AI-focused social posts.

The social-media finding also exposes a measurement problem. Logged-in feeds, dynamically loaded posts, app-only sessions, and algorithmic recommendation streams were often unavailable to the researchers. The visible sidebar could be measured more reliably than the personalized feed that held most of the user’s attention. In this category, the study may overrepresent stable interface elements and underrepresent actual content exposure.

Across all three settings, AI appears online as part of the interface rather than as a separate destination. That makes brand language, labeling, source visibility, and user controls more consequential. It also means any estimate of AI use that counts only chatbot sessions will miss a large share of AI-mediated information and commerce.

The Method Is Strongest as a Map of Exposure, Not a Census of Attention

Pew purchased March 2025 browsing data from 900 U.S. adults in KnowledgePanel Digital. Members had installed the RealityMeter application on qualifying personal devices and agreed to share browsing activity. Ipsos’ broader KnowledgePanel uses probability-based recruitment and address-based sampling, and Pew weighted the study to population benchmarks. The overall sampling margin of error was plus or minus 4.9 percentage points, with larger margins for age subgroups. Pew’s methodology explains the recruitment, weighting, collection, scraping, and classification procedures.

The researchers received URL logs with device, time, and visit-duration data. They removed zero-second visits and combined near-duplicate records. Pages associated with known malware, adult content, and selected productivity services requiring login were excluded from scraping. Failed pages and most login-protected or dynamically loaded content remained in total visit counts but could not be checked for AI terms.

Of 1,107,424 distinct URLs represented in the analysis, the team retrieved page content for 965,136. Collection occurred during April 7-17, 2025, after the March browsing period. A page might have changed between the visit and retrieval date, and some content could not be reconstructed. Images, video, audio, JavaScript-loaded material, paywalled text, and personalized feeds could be missed. Google results required a third-party scraping service because conventional retrieval methods do not reliably capture those pages.

Keyword matching supplied the broad exposure measure. Pew assembled a list ranging from generic terms such as “AI” and “algorithm” to product names, companies, model families, and technical concepts. The list sought specificity, yet any keyword approach can encounter ambiguity. “Algorithm” can appear in many settings, and a brand name can sit in navigation text rather than the material a user reads.

A logistic regression classifier separated substantive pages from incidental mentions. Human coders labeled 509 training pages and 400 evaluation pages. Agreement was high, with Cohen’s kappa values of 0.877 and 0.837. The model achieved an F1 score of 0.829, recall of 0.970, and precision of 0.724. High recall means it captured most substantive pages in the evaluation set. Lower precision means some pages classified as substantive were false positives. The 49% respondent-level figure should be read with that classification uncertainty in mind.

The sample and collection choices also set boundaries around generalization. The findings describe U.S. adults who qualified for the panel, used tracked personal devices, and browsed during one month. Work devices, untracked devices, native applications, private messages, voice assistants, smart televisions, and offline encounters fall outside the measure. The study remains valuable because those limits are documented and because observed behavior answers a different question from a survey. It does not measure every AI encounter. It shows where measurable web exposure occurred and how sparse deliberate engagement remained within a large volume of browsing.

AI Online Exposure Expanded After the 2025 Measurement Window

The 2025 browsing study captured a transitional month. Google’s AI summaries were already reaching most participants, but direct visits to generative AI tool websites remained limited to 13%. During 2025 and the opening seven months of 2026, search services, browsers, office software, shopping platforms, phones, and media products added more generated answers and assistants. The boundary between “using AI” and “using a service that contains AI” became harder to draw.

Pew’s February 2026 survey found that 49% of U.S. adults reported ever using AI chatbots, compared with 33% in 2024. Twenty-four percent said they used chatbots at least daily, and 44% reported using ChatGPT. The questions and respondent routing changed between some survey years, so the percentages should be interpreted with the qualifications supplied in the Pew report.

News use shows the same split between visibility and habit. The Reuters Institute Digital News Report 2026 found that 10% of respondents across its surveyed markets used AI chatbots for news each week, up from 7% in 2025. Usage reached 16% among respondents under age 35. Only 1% described AI as their main source of news, indicating that chatbot-based news use remained supplementary for most users.

Search products also became more conversational. In January 2026, Google added a smoother path for asking follow-up questions from an AI Overview and continuing in AI Mode. Its May 2026 Search update described expanded personalization, AI-assisted shopping, generated interfaces, agent-like functions, and wider use of newer Gemini models.

Those product changes do not erase the 2025 findings. They change the questions researchers need to ask. A future browsing study would need to distinguish exposure, reading, interaction, source opening, follow-up questioning, task completion, and later recall. It would also need application-level data because much chatbot and social activity occurs outside traditional web pages.

Longitudinal claims require caution. The 2025 tracking study, 2026 Pew survey, Reuters Institute survey, and Google usage figures use different populations, definitions, and methods. They cannot be combined into a single growth rate. Their shared direction is more defensible: generated answers and chatbot interfaces became more visible, but depth of use, source checking, and news reliance remained uneven.

Publishers, Brands, and Public Institutions Face a New Measurement Problem

Publishers once treated a search impression as a step toward a page visit. AI-generated search can break that sequence. A source may influence an answer without receiving a click, subscription, advertisement view, donation, or direct relationship with the person who benefits from the information. New Space Economy’s coverage of open-web traffic and Google’s business model describes the commercial tension between better in-platform answers and reduced referral value.

The response should not be reduced to producing more pages. Search systems and users have less need for generic summaries that can be recreated from common sources. Publishers gain stronger reasons to invest in original reporting, named expertise, direct access to documents, proprietary data, clear charts, current corrections, and explanations that connect events over time. These formats give a person a reason to open the source and give retrieval systems clearer evidence to attribute.

Direct audience channels become more important under lower click-through conditions. Email newsletters, memberships, browser notifications, RSS feeds, podcasts, events, and recognizable brands reduce dependence on a single discovery platform. Success measures also need to expand beyond rankings. Branded searches, returning visits, newsletter growth, direct traffic, engagement with original material, and revenue per visit can reveal whether a publication is building recognition even as generic search clicks fall.

Google’s AI features guidance says ordinary search requirements still apply to AI Overviews and AI Mode. A page must be indexed and eligible to display a search snippet, and Google does not require special technical markup for inclusion. Google’s newer generative search guidance recommends useful original content, crawlable text, accurate structured data, effective internal links, and sound page experience. It also states that Google Search does not use special files such as llms.txt to determine visibility in its generative search features. These documents describe Google’s own systems and do not guarantee indexing, placement, traffic, or inclusion in a generated response.

Brands face a different problem. The Pew data shows that marketing language exposed many people to AI labels, yet those labels often said little about function or evidence. Companies should name what the feature does, identify meaningful limits, disclose when text or images are generated, and avoid treating “AI-powered” as a substitute for a product description. Clear claims help consumers compare products and help journalists or regulators determine whether advertised capabilities match performance.

Public institutions carry a source-quality responsibility. Agencies, universities, standards bodies, and research organizations can publish stable pages with visible dates, clear authorship, downloadable data, accessible text, and revision histories. Generated answers are more likely to be useful when authoritative material is easy to retrieve and interpret. Weak source presentation leaves room for outdated copies, promotional summaries, or unsupported claims to fill the gap.

The wider information risk is not confined to false outputs. AI risks in 2026 include source loss, automated repetition, synthetic persuasion, privacy failures, cybersecurity weaknesses, and low-quality material crowding reliable work. Pew’s study adds a behavioral warning: users can encounter AI repeatedly without seeking it, understanding it, or opening the sources behind it. Measurement systems need to count mediation as well as adoption.

Summary

Pew’s browsing data shows a web in which artificial intelligence had already become ordinary interface material by March 2025. Most participants reached a page containing an AI term, more than half encountered a Google-generated search summary, and roughly half visited at least one page classified as substantively related to AI. Yet sustained attention was scarce relative to the scale of total browsing. The typical substantive visitor reached only three such pages during the month.

That pattern places a limit on simple adoption statistics. A person may receive an AI-generated answer, read a review summary, see an AI product label, or pass a chatbot link without making a deliberate choice to use artificial intelligence. Counting only direct chatbot visits understates mediated exposure. Counting every keyword match as active use overstates engagement.

The study’s limitations are part of its meaning. Browser tracking can observe pages but not reliably determine attention, comprehension, trust, or downstream decisions. Missing application activity and personalized feeds leave significant areas of digital life unmeasured. The substantive-page classifier performed well but produced false positives. The one-month U.S. sample cannot stand in for every country or every stage of AI adoption.

Research published through August 2, 2026, indicates that generated search summaries and chatbot-based news use continued to spread. Company announcements describe larger search integrations, and independent surveys show broader reported use. Yet source opening, sustained engagement, and regular news reliance remained lower than exposure. The information system has moved faster than the public’s ability to identify where generated material begins, which sources support it, and when a direct source visit is needed.

The next measurement standard should connect interface exposure with human response. It should record whether a person expanded a summary, opened a source, asked a follow-up, compared accounts, completed a purchase, changed a belief, or returned to the same publisher. Page visits alone cannot answer those questions. Survey recall alone cannot reconstruct the interface that shaped the choice.

AI online is no longer best understood as a collection of separate tools. It is a layer inserted into search, news discovery, commerce, social interfaces, and routine software. The practical issue is not whether people encounter it. The issue is how often that mediation replaces direct contact with evidence, authors, institutions, and original reporting.