Brands now ask a question that barely existed three years ago: when someone puts a question to ChatGPT, Gemini, Google's AI Overviews or AI Mode, Perplexity, Copilot or Claude, does the answer mention us, cite us, or send anyone our way? The evidence for answering it is growing quickly, but it is uneven. Most of the large datasets come from companies that sell search or AI-visibility software and do not publish their raw data. The academic work is more transparent, but smaller, and often run against model APIs rather than the consumer products people actually use.

This report brings together what the published record shows as of September 2026. Every figure below is attributed to its source, with the sample and period that produced it. Where studies disagree, we say so rather than averaging them.

AI answers are now a default search surface

Scale is the least disputed part of the picture. In Alphabet's second-quarter earnings remarks on July 22, 2026, Sundar Pichai said AI Mode had passed 1 billion monthly active users since its global expansion the previous October, and that the Gemini app had 950 million monthly active users.

How often an AI answer appears on a given search is harder to pin down, because every tracker samples different queries. In Pew Research Center's analysis of browsing data from 900 US adults, 18% of 68,879 Google searches made in March 2025 produced an AI summary. Keyword-tracking firms report a wide range:

  • Semrush found AI Overviews on 15.69% of its 10 million-plus tracked keywords in November 2025, down from 24.61% in July.
  • Conductor found them on 25.11% of 21.9 million US searches between September 15 and October 12, 2025.
  • BrightEdge said they appeared on roughly 48% of its tracked keywords in February 2026, up from about 31% a year earlier.

These figures are less contradictory than incomparable. A keyword set weighted toward long informational questions will show more AI answers than one weighted toward product and navigational terms. Pew saw the same pattern in real behavior: AI summaries appeared on 8% of one- or two-word searches and 53% of searches of ten words or more.

Beyond Google, adoption for news remains modest. The Reuters Institute's 2026 Digital News Report found weekly use of AI chatbots for news rose from 7% to 10% across its 48 markets, with no increase in the US, UK, France or Germany.

Referral traffic: steep growth from a small base

The major datasets agree on direction and disagree on size.

  • Similarweb's 2026 Generative AI Landscape data counts an average of 770.7 million AI referral visits a month worldwide between June 2025 and May 2026, up 117.4% year on year. In a separate mid-2025 analysis, Similarweb found ChatGPT accounted for more than 80% of AI referrals to the top 1,000 global domains. It notes that visits arriving without referrer data land in direct traffic, so its counts are a floor.
  • Conductor's benchmarks, drawn from 1,215 enterprise customer domains in the US between May and September 2025, put AI referrals at 1.08% of all website visits, with ChatGPT responsible for 87.4% of them and growth of roughly 1% a month.
  • Adobe, analyzing more than 1 trillion visits to US retail sites, reported via TechCrunch that AI traffic rose 393% year on year in the first quarter of 2026 and converted 42% better than non-AI traffic in March 2026, reversing a gap a year earlier, when AI visitors converted 38% worse.

Product decisions can move these numbers overnight. Similarweb data reported by Search Engine Roundtable showed ChatGPT referral visits rising about 150% in the week after May 7, 2026, when ChatGPT began showing more prominent brand links, with around 60% of that traffic landing on homepages.

The other side of the ledger is crawling. Cloudflare's 2025 Radar Year in Review compared how often each platform's bots fetch pages with how often its products send visitors back. After an erratic first five months, Anthropic's ratio ranged from about 25,000 to 100,000 crawls per referral; OpenAI's peaked near 3,700:1 in March 2025; Perplexity's stayed below 200:1 from September; Google's ranged between roughly 3:1 and 30:1. Cloudflare cautions that native apps often send no referrer header, which inflates these ratios by an unknown amount.

Clicks from Google's AI answers: the most contested number

Studies of what AI Overviews do to clicks measure different things and reach different conclusions.

  • Pew's panel showed users clicked a traditional result on 8% of visits to pages with an AI summary, against 15% without one, and clicked a link inside the summary on 1%.
  • Ahrefs, comparing aggregated desktop Search Console data for 300,000 keywords in December 2025 with December 2023, associated AI Overviews with a 58% lower click-through rate for the top-ranking page.
  • Seer Interactive, tracking 5.47 million queries for 53 brands, found organic CTR on AI Overview queries fell to 1.3% in December 2025, then recovered to 2.4% by February 2026. Pages cited in the overview earned 120% more organic clicks per impression than uncited ones, yet still 38% fewer than on queries with no overview.
  • Semrush and Datos, its clickstream company, found that when the same keywords gained an AI Overview, their zero-click rate fell slightly, from 33.75% to 31.53%.
  • Google's Liz Reid wrote in August 2025 that total organic click volume from Search was relatively stable year on year, without publishing supporting data.

Click-through rate per query, zero-click rate and total click volume are different quantities; one can fall while another holds if search volume grows. None of the underlying datasets is public.

What is known about how sources get chosen

Google documents the most. Its guide to generative AI features, updated in July 2026, says AI Overviews and AI Mode ground answers in pages retrieved by its core ranking systems and issue "query fan-out" searches on related subtopics. A page must be indexed and eligible for a snippet; no special files or schema are required. Since August 31, 2026, site owners worldwide can also exclude their site from these features, which Google says does not affect ranking elsewhere in Search.

Fan-out may explain why citations increasingly come from beyond page one. Ahrefs found 76% of AI Overview citations ranked in the top 10 in July 2025 and about 38% in March 2026, although it changed its parsing method in between. BrightEdge puts top-10 overlap at about 17%.

The best-known experiment is the GEO paper by researchers from Princeton and IIT Delhi, presented at KDD 2024. Adding citations, quotations and statistics to source text raised visibility by up to 40% in a simulated engine built on GPT-3.5 over a 10,000-query benchmark; keyword stuffing did not help. It is a lab result on 2023-era models, and its real-world check used 200 queries on Perplexity with sources supplied as file uploads.

Observational studies point to off-site reputation. Across 75,000 brands, Ahrefs found brand web mentions correlated with AI Overview visibility at 0.664 (Spearman), against 0.218 for backlinks, and cautions that correlation is not causation. Across 17 million citations, Ahrefs also found the pages AI assistants cited were on average 25.7% younger than those in organic results, with AI Overviews the exception. We cover the mechanisms in depth in how AI answer engines choose which sources to cite.

Engines differ, and so do the studies of them

Profound's analysis of 680 million citations from August 2024 to June 2025 found Wikipedia accounted for 7.8% of all ChatGPT citations, while Reddit accounted for 6.6% of Perplexity's and 2.2% of AI Overviews'. Those shares move: Semrush tracked ChatGPT citing Reddit in close to 60% of responses in early August 2025 and around 10% by mid-September. Ahrefs found only 12% of URLs cited by ChatGPT, Gemini, Copilot and Perplexity ranked in Google's top 10 for the same prompt; Perplexity stood out at 28.6%.

Studies also disagree about which kind of source wins. A University of Toronto preprint found AI engines strongly favored earned media over brand-owned and social content in product-ranking queries, and Muck Rack attributes 84% of citations in its dataset to earned media. Yext, studying location-based queries in four industries, attributes 86% of 6.8 million citations to websites and listings that brands manage. The answer depends on the questions asked; our review of cross-engine comparisons and our industry data roundup examine why.

Accuracy remains a caveat for every engine. The EBU and BBC study of 2,709 responses from the free versions of ChatGPT, Copilot, Gemini and Perplexity, generated in late May and early June 2025, found 45% had at least one significant issue and 31% had serious sourcing problems.

Open questions

  • How often do real users' prompts trigger web retrieval on each engine? Most published figures come from test prompts, not usage logs.
  • How many clicks do AI answers actually send? Google's new Search Console report for generative AI features shows impressions, and Bing's AI Performance report counts citations but not clicks.
  • Which signals cause citation, rather than merely correlate with it? Nearly all industry evidence is observational.
  • How stable is any single measurement? SparkToro found fewer than one in 100 repeated runs returned the same list of brand recommendations, and a 2026 preprint by Schulte, Bleeker and Kaufmann argues visibility should be reported as a distribution.
  • How much AI-driven traffic is invisible because apps strip referrers?
  • Do findings from English-language, US-centric datasets hold elsewhere?

For practitioners weighing what to change, our comparison of AEO and SEO sets Google's guidance against the evidence.

What we're measuring next

TheAEOWire is preparing original studies aimed at these gaps: repeated sampling of the same prompts across engines so that citation share can be reported with its variance, a strict separation of mentions, citations and clicks, and full disclosure of prompt sets and collection dates. We will publish methods before results and will not report a number until the data behind it exists. The plan is described in our study program and on our methodology page.