AI Search Statistics (2026): Usage, AI Overviews, Citations

AI assistants are huge but still a small share of search behavior, Google's own AI Overviews may be the biggest AI search surface, and AI citations are often wrong or unrelated to Google rankings. OpenAI says ChatGPT has 1.2 billion weekly users, yet SparkToro's panel puts all AI tools at 3.2% of U.S. desktop searches.

Updated 3 October 2026. We re-checked every figure below against its source on that date, and replaced twenty thinner statistics pages with this one (part of our merge of 165 pages into 42). Each group ends with what it means for your brand.

How many people use AI search and assistants?

Hundreds of millions of people use AI assistants every week, but "share" depends on the yardstick. Companies report weekly or monthly users, Pew asks adults, and trackers count web visits. These figures are not interchangeable.

Two readings stand out. Statcounter's 80.5% and Sensor Tower's "below 50%" describe the same chatbot with different denominators, so never quote one without the other. And since Semrush's panel saw web search on about a third of ChatGPT prompts, most answers still draw on memory, not live retrieval.

What it means for your brand: assistants are mainstream enough to matter, and Claude's 1.0% in a web-analytics tracker cannot see API or embedded use. Find out which engine your own buyers use before you optimise for any single one.

How big are Google AI Overviews and AI Mode, and what happens to clicks?

SparkToro calls Google the largest AI search tool, because AI Overviews appear on roughly 16% of its results pages, and the click evidence conflicts. Pew's browsing panel and Ahrefs see large losses; Semrush's same-keyword test sees none.

Why the conflict? Pew watched real people. Ahrefs compared two keyword samples two years apart. Semrush followed the same keywords before and after an AI Overview appeared and found slightly more clicks, noting that AI Overview keywords were already high in zero-click behavior.

Our position: the damage is real for simple informational queries and uneven elsewhere, so one headline percentage is the wrong tool. Our AI Overviews and your traffic guide covers what to do, and is SEO dead covers why "SEO is dead" is only half right. A zero-click search is a Google-side metric, not automatically your loss.

What it means for your brand: open Search Console, filter to informational queries that trigger AI Overviews, and compare click-through before and after. That is your number; the studies above are only the range.

How accurate are AI citations?

A visible citation is a pointer, not proof. Depending on the test, between a quarter and well over half of AI search citations fail to support the claim or identify the right source. Measure link validity, source identity and claim support separately.

The first two rows are not contradictory. The Tow Center asked tools to name a known article; Liu and colleagues checked whether cited pages support generated sentences. That is why there is no universal hallucination rate: NewsGuard's chatbots answered every prompt in 2025, up from refusing many a year earlier, and 35% of answers were wrong.

Rankings also predict little. Ahrefs found 76% of AI Overview citations come from top-10 pages, while assistants mostly do not follow Google's results. Semrush's Reddit swing shows how fast citation sources move.

What it means for your brand: being cited is not being described correctly. Save the answer and the cited page, and re-run the same questions monthly. Our crawler directory shows which bots feed which engine.

Do people trust AI answers when they shop?

Trust runs ahead of accuracy, and shopping traffic is converting better than a year ago. In a randomized experiment, citations raised trust even when the links were wrong, while Adobe now measures AI-referred shoppers converting above other traffic.

A year ago the same Adobe series showed AI visitors converting worse. That reversal is the strongest argument against old statistics pages, including our own earlier version of this one.

What it means for your brand: because shoppers trust a confident answer, a wrong price or feature in an AI answer costs more than an absent one. Check what engines currently say about you, and make product pages readable without scripts.

What does AEOeye's own data show?

Our data says brands rarely block AI search bots, publishers often do, and a brand's real problem is absence from the sources AI reads. Both datasets below are ours, small and self-selected, and we state their limits.

In AEOeye's crawler access study (3 October 2026, 384 sites with readable robots.txt, a convenience sample of well-known brands; per-site CSV):

• 4.2% of sites (16) block OAI-SearchBot, ChatGPT search's crawler; 9.1% (35) block GPTBot.
• 1 of 193 SaaS sites blocks any AI retrieval bot, versus 27 of 38 news publishers.
• 60% of SaaS sites we could check publish an llms.txt, against 1 of 29 publishers. We still think llms.txt is overrated as a ranking lever.
• 16% of 375 homepages challenged our self-identified crawler, including 27.3% of e-commerce sites.

In our recommendation benchmark (127 completed audits, 27 brand names, 892 answers, June to September 2026), brands were named in 17.6% of answers. When named and ranked, 71.4% of 147 mentions came first, and answers named 2.9 competitors on average.

The caveat is mandatory: these are self-selected brands, engine samples are unequal and not comparable, and four engines had under 10 mentions. Do not rank engines from them. See how to measure AI brand recommendations, or run the free AEOeye audit on your own buyer questions.

How should you read AI statistics without being fooled?

Ask four questions of every number: what unit, what yardstick, what design, and what date. Most bad AI statistics fail one of these, usually because a definition was dropped on the way to the headline.

• Unit. ChatGPT's weekly users and Gemini's monthly users are different clocks; Perplexity reports queries. Companies audit their own counts, and clickstream panels see only their members.
• Yardstick. Statcounter counts search referrals and puts Google near 90%; SparkToro's 41-site desktop panel counts searches and says 73.7%. Both are right about different things.
• Design. Surveys record what people say, panels record what they did, and forecasts record neither. Ahrefs' 58% is a correlation, not a measured cause.
• Date. Citation sources swung within weeks, and AI Overview coverage rose to 24.61% then fell to 15.69% inside one year.

Match the study to the claim:

• Does a content tactic raise visibility? The GEO paper, a controlled benchmark.
• Do people click less? Pew for observed clicks, with the caveats above.
• Are citations correct? ALCE and the Tow Center.
• Are answers stale? FreshQA, where all tested models struggled with fast-changing facts.

Benchmarks measure a task, not the market. Pin the version and date, and keep one stable test set plus one private holdout, so a score gain is not just a leaked question.