AI Search & GEO Statistics 2026: What 16 Verified Sources Say About the Shift

Two months ago I wrote a GEO playbook documenting how MacMD Viewer went from 6 to 130 ChatGPT referrals per month after applying a specific framework. That post is the practitioner side — what to do, line by line. This post is the macro side: how big is the AI-search shift really, how fast is it moving, and what does the verified data actually say.

The setup is the same as my AI productivity stats post: I dispatched research agents, they came back with dozens of stats, and most of them were wrong. Not "directionally fine" wrong. Specifically wrong — Cloudflare crawl-to-refer ratios inflated, BrightEdge percentages invented, llms.txt adoption rates off by an order of magnitude. The same vendor-blog telephone game is happening across the AI search space, just with bigger stakes because nobody has been measuring this category long enough to have stable benchmarks.

So this is a deliberately shorter list. 16 stats, every one verified by opening the source URL on the day this was published and confirming the number is exactly what's on the page.

TL;DR — what the 2026 AI search data actually says:

1. AI Overviews now appear on more than half of Google queries, and coverage grew 58% year-over-year between Feb 2025 and Feb 2026 (BrightEdge, via Search Engine Journal). Some sectors are well past that — healthcare 88%, education 83%, B2B tech 82%.
2. AI Overviews cut organic clicks by 38% in the first randomized field experiment on the subject (Agarwal & Sen, Indian School of Business + Carnegie Mellon, n=1,065 Chrome users, SSRN April 2026). That's the cleanest causal estimate available.
3. Being cited inside an AIO matters more than ranking #1 under it. Seer Interactive's full-year 2025 study finds being cited delivers +120% more organic clicks vs not being cited on the same SERP.
4. AI crawlers take far more than they give. Anthropic's ClaudeBot crawls 50,000 pages for every referral it sends back. OpenAI's GPTBot: 887:1. Perplexity, for News & Publications, runs about 88:1 — by far the most efficient of the three (Cloudflare).
5. The infrastructure standard nobody adopted (yet). Only 0.3% of the top 1,000 websites have an llms.txt file as of June 2025 (Rankability). The first-mover window is unusually wide.

Why This List Is Shorter Than the Others You'll Find

The AI search category is roughly two years old as a measurable thing. There's not enough stable methodology yet to support 80-stat roundups. Most of the ones you'll find online are stitching together numbers that don't agree, from sources that lifted them from other roundups, with citation chains that go cold after one or two hops.

I started with 75+ candidate stats from two AI research agents. I threw out everything I couldn't trace back to one of these primary sources:

• Platform first-party announcements: OpenAI (via TechCrunch), Anthropic, Google
• Cloudflare Radar — their AI Insights data is the only consistently-measured AI crawler dataset that exists
• SEO-industry research with stated methodology: BrightEdge (via Search Engine Journal), Seer Interactive R&D
• Academic / peer-reviewed: Aggarwal et al. (KDD 2024), Agarwal & Sen (SSRN April 2026)
• Analyst firms: Gartner
• Direct measurement researchers: Rankability (llms.txt adoption tracker)

What survived: 16 stats, 8 unique source domains, every URL fetched on 2026-05-26. The aggregator blogs that fed most of the dropped stats are listed at the bottom under "Sources Discarded" if you want to verify why.

ChatGPT, Claude, Gemini, Perplexity: Who's Actually Big

The headline number you need first: ChatGPT had 900 million weekly active users by late February 2026, up from 800 million in October 2025 (per OpenAI's own announcement, via TechCrunch). OpenAI also reports 50 million paying subscribers as of the same period. For context, that's roughly 1 in 8 people on the internet using ChatGPT each week.

I'd love to give you matching numbers for Claude, Gemini, and Perplexity, and the research agents tried — but most of the specific Anthropic / Google / Perplexity user-count figures I could find traced back to vendor SEO blogs paraphrasing partial information or projecting from outdated press releases. Until each company publishes consistent definitions of "weekly active" or "monthly query volume," any precise number you see for non-ChatGPT platforms is best treated as directional rather than precise. The relative pattern that does appear consistently across the credible sources: ChatGPT remains dominant, Gemini has been growing fastest among the four, and Perplexity is by far the smallest but punches above its weight in referral efficiency (see the crawler section below).

AI Overviews: They're On More Than Half of Queries Already

BrightEdge has been tracking AI Overview coverage longer than anyone else with credible methodology. Their year-over-year comparison, reported by Search Engine Journal in March 2026, is the cleanest read available on how fast Google's AI summaries are spreading.

Sector	Earlier 2025 share triggering AIO	December 2025 share
Healthcare	~72% (since 2024)	88%
Education	18% (May 2025)	83%
B2B technology	36%	82%
Restaurants	10%	78%

Overall AIO coverage grew 58% year-over-year between Feb 2025 and Feb 2026. The healthcare number is the one that should land hardest if you're a marketer — Google has been showing AI Overviews on roughly 72% of healthcare queries since 2024, and by late 2025 that rose to 88%. Healthcare was first to get the AIO treatment because the queries are informational and the answers come from authoritative sources Google can defer to. Other sectors are catching up fast, with education's 65-point jump in seven months the most dramatic move on the chart.

The takeaway from this table isn't "AI Overviews are coming" — they're here. The question for 2026 has already shifted to "what happens to traffic when they show up."

The Click Math: 38% Fewer Outbound Clicks, But Citation Beats Ranking

Until April 2026, every estimate of the AI Overview traffic impact came from observational studies — site owners watching their Google Search Console numbers compress and trying to attribute. Useful, but not causal. That changed when Agarwal & Sen at the Indian School of Business and Carnegie Mellon published the first randomized field experiment on the topic (SSRN working paper, April 2026, abstract ID 6513059).

Their setup: 1,065 US Chrome users, randomly assigned to three groups for two weeks. Control saw Google normally. "Hide AIO" had a Chrome extension remove AI Overviews in real time. "AI Mode" was redirected to Google's AI-first search experience. Then they compared everyone's click behavior on identical query types.

The headline finding: AI Overviews reduce organic clicks to external websites by 38% on queries where they appear. Self-reported search satisfaction stays nearly unchanged when the AIOs are removed — users get what they need from the AI summary without clicking, and they don't seem to mind. The causal direction is now established: AIOs are responsible for a meaningful chunk of the click compression marketers have been observing for the past year.

Seer Interactive's parallel work, also published in April 2026, gives a more granular view. Their full-year 2025 study tracked 53 brands across 5.47 million queries and 2.43 billion organic impressions. Two findings worth pulling out:

• Brand-cited organic CTR dropped 61% between Q3 and Q4 2025 — falling from 2.52% in September to 0.76% in November. But clicks stayed essentially flat (398K → 400K → 302K). Impressions more than doubled in the same period (15.8M → 33.1M → 39.5M). The compression came from the denominator growing far faster than the numerator, not from raw click loss.
• Being cited delivers +120% more organic clicks vs not being cited on the same SERP. Throughout 2025, cited domains kept 2-5× the CTR of non-cited domains, even as both compressed. On transactional queries, the gap is starker: not-cited organic CTR fell to 2.15% from 4.17% — a 48% decline across 1.73 million transactional impressions.

The takeaway is counterintuitive: AI Overviews aren't bad for traffic if you get cited inside them. They're devastating for everyone else. The competitive question has shifted from "rank #1" to "be the source the AI quotes."

What Actually Earns Citations: The Academic Anchor

The foundational research on what increases generative-engine citation is Aggarwal et al., "GEO: Generative Engine Optimization" (arXiv 2311.09735, KDD 2024). Their benchmark — GEO-bench — runs 10,000 queries across the major generative engines and tests which content techniques increase visibility in AI responses.

The headline result: up to 40% visibility lift in generative engine responses from three main techniques: citation inclusion (linking to sources), quotation addition (using direct quotes), and authoritative language. That paper is the only peer-reviewed benchmark in the space as of mid-2026, and it's still cited as the methodological gold standard.

The community evidence layered on top of that — including the structural-pattern studies and author-bio correlation work I documented in my GEO playbook — extends the same direction. Lists get cited 3× more than paragraphs. Author bios correlate with +47% more citations. Structured comparison tables outperform prose for AI extraction. The pattern is consistent enough across independent measurements that the playbook is actually settled: optimize for the AI to find, extract, and quote you, and you'll get cited. The thing that's still being measured is how much lift each tactic delivers in different categories.

If you want the tactical version with code samples and a real domain's before/after data, the GEO playbook walks through everything I changed on macmdviewer.com to go from 6 to 130 ChatGPT referrals per month.

AI Crawlers Are Taking Far More Than They Give

This is the data that most people building for the open web haven't fully internalized. Cloudflare publishes the only consistent dataset on AI crawler activity vs referral traffic — their crawl-to-refer ratio metric tracks, for each platform, how many pages it crawls vs how many human visitors it sends back to your site.

The numbers from Cloudflare Radar (default view across their full customer base, August 2025):

Crawler	Crawl-to-refer ratio (default view)	News & Publications
Anthropic / ClaudeBot	~50,000:1	8,800:1
OpenAI / GPTBot	887:1	401.7:1
Perplexity	(not separately tracked default)	88:1

Industry filter shown is Computer and Electronics — these were the ratios Cloudflare published for that vertical. News and Publications ratios are different again; see the source post for the full breakdown.

Read those numbers slowly. Anthropic's ClaudeBot crawls fifty thousand pages for every one human visitor it refers back to a publisher. OpenAI's crawler is in the same direction at 887:1, even at its most generous ratio in the News & Publications industry filter. Perplexity is the outlier on the efficiency side — for News & Publications, they're at 88:1, an order of magnitude better than OpenAI and 100× better than Anthropic.

The narrow practical implication: if you're a small site, ClaudeBot is taking your content and converting almost none of it into traffic for you. The broader strategic implication: the search-traffic deal that powered the open web for 25 years (we crawl, you rank, we send eyeballs) is being unilaterally renegotiated by the AI platforms in real time, and the new contract is far worse for publishers. This is why so many news sites are blocking GPTBot and ClaudeBot in robots.txt — and why others (like me) are choosing to allow them and bet on citation referral instead.

Both bets are reasonable. The one you don't want to make is the unconsidered one.

llms.txt: The Standard Almost No One Has Adopted

llms.txt is the proposed standard for telling AI crawlers which pages on your site are most important and how they should be cited (similar to how robots.txt works for traditional search bots, but designed for the AI era). It's been around since mid-2024, has a published spec at llmstxt.org, and was a key part of my own GEO setup.

How widely has it been adopted? 0.3% of the top 1,000 websites globally as of June 2025, according to Rankability's automated scan — that's 3 sites out of 1,000. The major platforms (Google, Facebook, Amazon, Microsoft) all have nothing.

There are two ways to read that number. The pessimist's read: nobody cares about llms.txt yet, the standard is dead on arrival. The optimist's read: the first-mover window is unusually wide. If you adopt now, you're competing with a tiny set of other early adopters for whatever AI-platform attention the standard ends up being worth. I bet on the optimist read.

The Gartner Forecast: 25% Search Decline by End of 2026

The longest-standing macro forecast for the AI-search shift comes from Gartner's February 19, 2024 press release: traditional search engine volume will drop 25% by 2026, with search marketing losing market share to AI chatbots and other virtual agents. That was an aggressive call when it came out — Google had just launched Bard, AI Overviews wasn't yet rolled out — and it's the prediction the rest of the industry has been measuring against.

We're now in mid-2026, the deadline year. Based on the BrightEdge AIO coverage data (above) and the Agarwal & Sen 38% click reduction finding, the trajectory matches: outbound clicks from Google to publisher sites have compressed substantially, even if Google's total query count hasn't fallen the predicted 25%. Gartner's forecast looks directionally right; whether it lands precisely depends on how you define "search volume" (queries vs clicks vs traffic) and which platforms you count.

What This Actually Means If You're Marketing or Building in 2026

Five honest takeaways from sitting with this data:

The era of "rank #1 on Google" is structurally over for most informational queries. Above the fold of a 2026 search result is an AI Overview, not your blue link. The new objective is being one of the 3-5 sources the AI Overview quotes — that's where the +120% click delta lives.

The platforms haven't settled their referral behavior yet. Anthropic crawls 50,000 pages per referral; Perplexity is at 88:1 for News & Publications. There's no rational economic reason this stays the way it is — Anthropic has every incentive to behave more like Perplexity over time as the model providers realize crawl-without-refer isn't sustainable. Watch this number. If ClaudeBot's ratio doesn't improve by end of 2026, the publisher community will keep escalating the blocking strategy and the AI platforms will lose access to fresh, high-quality content. Game theory says they'll move.

The Aggarwal et al. paper is still the only peer-reviewed benchmark in this space, and the +120% citation premium from Seer Interactive is the strongest before-and-after data outside of academia. If you're making a 2026 content strategy and you're not designing for AI citation pickup, you're optimizing for a smaller-and-shrinking surface area.

The infrastructure plays are unusually cheap right now. llms.txt adoption at 0.3% means implementing the standard takes you from the long tail to the leading edge in 15 minutes. Schema markup, FAQ structure, structured tables — all known SEO tactics that now also serve AI citation pickup. The cost of the tactical work hasn't changed; the payoff has.

The smartest people I know are running an A/B portfolio: block on critical content (paywalled investigative journalism), allow on top-of-funnel (educational content where citation pickup is the whole point), and measure the trade-off. There is no single right answer for everyone — but there is a clearly wrong move, which is to let your default stay "block everything" or "allow everything" without a model of which content type fits which strategy.

I build MacMD Viewer precisely because of where these trends are going: AI-native workflows produce vast amounts of markdown that needs to be read natively on a Mac, and the people writing about that workflow (the journalists who'd cite a stats page like this one) are also operating in this new world. The data here is part of why I'm bullish on building for the AI-tooling-tailwind side of the macOS market.

FAQ

How many AI Overviews appear in Google search results in 2026?

Coverage grew 58% year-over-year between February 2025 and February 2026 according to BrightEdge research, with significant industry variation — 88% in healthcare, 83% in education, 82% in B2B tech, and 78% in restaurants by late 2025.

How much do AI Overviews reduce organic clicks?

A randomized field experiment by researchers at the Indian School of Business and Carnegie Mellon (Agarwal & Sen, SSRN April 2026, n=1,065 US Chrome users) found AI Overviews reduce organic clicks to external sites by 38% on queries where they appear, while self-reported search satisfaction stays nearly unchanged when removed.

What's the click-through rate impact of AI Overviews on cited brands?

Seer Interactive's full-year 2025 study (53 brands, 5.47M tracked queries, 2.43B impressions) found brand-cited organic CTR dropped 61% from Q3 to Q4 2025 — falling from 2.52% in September to 0.76% in November. But being cited still delivers +120% more clicks vs not being cited on the same SERP.

How many people use ChatGPT in 2026?

ChatGPT reached 900 million weekly active users by February 27, 2026, up from 800 million in October 2025, according to OpenAI's own announcement. The company also reported 50 million paying subscribers.

How aggressive are AI crawlers vs the traffic they send back?

Cloudflare data shows Anthropic's ClaudeBot has a crawl-to-refer ratio of nearly 50,000:1 — it crawls 50,000 pages for every single referral it sends back. OpenAI's GPTBot is 887:1 (default view). For News & Publications specifically, the ratios are: Anthropic 8,800:1, OpenAI 401.7:1, Perplexity 88:1.

How widely has llms.txt been adopted?

Almost not at all. Rankability's June 2025 scan of the top 1,000 websites globally found only 0.3% — 3 out of 1,000 — had implemented an llms.txt file. The standard is still in early-adopter territory.

Sources

Every statistic above is sourced from one of the following primary organizations. URLs are in the audit trail at the bottom of the page (view-source).

• TechCrunch (citing OpenAI announcement) — ChatGPT 900M WAU + 50M paying subs (Feb 27, 2026)
• Search Engine Journal / BrightEdge — AI Overviews industry coverage data (March 2026)
• Agarwal & Sen — Indian School of Business + Carnegie Mellon (SSRN working paper, April 2026)
• Seer Interactive R&D — AIO Impact on Google CTR 2026 Update (April 2026)
• Aggarwal et al. — GEO: Generative Engine Optimization (KDD 2024, arXiv:2311.09735)
• Cloudflare Radar — AI Crawlers by Purpose and Industry (August 2025)
• Rankability — LLMS.txt Adoption Research Report (June 2025)
• Gartner — Press Release on Search Engine Volume Decline (February 2024)

Sources Discarded

For transparency, here are the secondary / aggregator domains where research agents pulled stats from that did not survive verification:

• relixir.ai, authoritytech.io, getpanto.ai, seo.com, stackmatix.com, ttms.com, technologychecker.io, frase.io blog, thestacc.com

In several cases the URL was real but the specific numbers attributed to it did not appear on the page — likely paraphrasing chains and round-up drift. If you want the original numbers I verified, the audit trail at the bottom of this page lists every source URL with the exact quote.

---

Fact-checked against primary source URLs on 2026-05-26. If you spot a number that's drifted from its source, email me at arthur@gettin.xyz — I'll fix it within the day and note the correction here.