What the numbers actually say about AI search in 2026

OpenAI’s engineering note of September 11, 2026, says Habitat, its application storage platform, now supports products used by more than one billion people each week. That is ChatGPT as a whole. It is not a count of people who search. Keep the label. In a March 31, 2026 funding note quoted by The Verge, OpenAI said search usage had nearly tripled in a year. That is the cleanest company sentence on Search as a behavior. It is a growth rate, not a headcount.

ChatGPT Search has been available to everyone in supported regions since February 5, 2025, including logged-out users. ChatGPT decides when to search, or the user taps the web-search control. Enterprise and Edu workspaces can turn the tool off. Pew’s February 17–23, 2026 survey of U.S. adults is the independent check on whether Americans treat chatbots as search. Those rates are “ever used,” not weekly actives. A 44 percent ever-use rate among U.S. adults can sit next to a one-billion-a-week global product. Different universes.

Shi, Zhu, and Gu, working from a household panel at Bocconi, found outbound referrals in 5.2 percent of ChatGPT conversation sessions against 31.1 percent of Google queries. Three-quarters of ChatGPT-active households in their window never clicked out. After wider access to ChatGPT Search, traditional search queries fell. Informational destinations took the larger referral cuts. The paper is a panel study, not a census. Cite the session rates. Do not turn them into a global traffic forecast. OpenAI Signals does not isolate Search.

On June 5, 2025, at Bloomberg’s Tech Summit, CEO Aravind Srinivas said Perplexity took 780 million queries in May. Growth was more than 20 percent month over month. He put daily volume at about 30 million and recalled 3,000 queries on day one in 2022. TechCrunch filed the remarks the same day. That is the last load-bearing usage number from the company. A 20 percent monthly rate does not survive a year of compounding. Do not write the 2026 query volume as if the May 2025 slope held.

What Perplexity is, for measurement: a cited answer on top of live retrieval. Academic comparisons treat Perplexity Sonar as the high-retrieval pole. In Characterizing Web Search in the Age of Generative AI, Sonar retrieved about 14 pages per query on average. GPT with search-as-a-tool retrieved less than one. Google AI Overviews and Gemini sat near nine. The same paper found generative systems citing domains outside the organic top 10. That is a source-selection finding, not a popularity ranking.

Pew’s U.S. adult chatbot list in February 2026 did not break out Perplexity. An academic consumer-AI survey put past-month Perplexity use among AI users at 10.1 percent and primary-assistant share at 0.5 percent. That sample is AI users, not all adults. Do not paste it next to Pew’s 44 percent without saying so. India is the one country where a third party published a usage path with a method: Sensor Tower via TechCrunch, after an Airtel free-Pro offer. Mobile-app activity in one country after a subsidy. Not global MAU.

Leave these in, labeled as embedded Search, not as rival apps. Google’s August 31, 2026 website-owner post puts AI Overviews above 2.5 billion monthly active users and AI Mode above one billion. Sundar Pichai’s Q2 2026 remarks repeat the AI Mode billion and add that AI features in Search send billions of clicks to websites each week. Those are exposure counts for surfaces inside Search, not logins to a separate app.

Overviews reached more than 200 countries and 40-plus languages by May 20, 2025. In the United States and India, Google said Overviews raised use of Google by more than 10 percent on the query types that trigger them. Preferred Sources arrived in May 2026. People had selected more than 345,000 unique sources. Google said they were twice as likely to click a Preferred Source.

Pew found 65 percent of U.S. adults at least sometimes see AI summaries in search results. Six in ten say they read them. On breaking news, 1 percent go first to an AI chatbot and 28 percent go first to a search engine. Query studies disagree on how often an Overview appears because they send different queries. Xu, Iqbal, and Montgomery found 13.7 percent overall activation and 64.7 percent on question-form queries. A separate benchmark put appearance at 51.5 percent. Report both. Do not average them. Of 98,020 atomic claims in the trending-query study, 11.0 percent were unsupported by the cited pages. That score is Google Overviews on trending queries. It is not a score for ChatGPT Search or Perplexity.