The State of AI Search in 2026: 42 Statistics, Every One Sourced

Most AI search statistics circulating in 2026 have no traceable source. A number appears in a blog post, gets quoted without attribution, gets quoted again with a different figure attached, and within two quarters it is being cited as established fact by people who have never seen the underlying study. What follows is 42 statistics where the producing organisation, the measurement period, and the method are all identifiable, grouped by what they measure.

TL;DR

• Every figure below names the organisation that produced it and the period it covers. Where the method is weak, the entry says so.
• The largest measured shift of the year is concentration collapse in AI referrals: ChatGPT fell from 89.1% to 62.6% of B2B AI referral share in Goodie's GA4 panel.
• The most-quoted number in the field, the Princeton study's 40% visibility gain, is a maximum rather than an average and applies only once a source is already in context.
• Crawler blocking is now a material variable: Cloudflare has blocked AI crawlers by default for new domains since July 2025, and changes its category defaults on September 15, 2026.
• A closing section lists widely circulated statistics that were excluded from this page because no primary source exists for them.

How to read this list

Three rules were applied. A statistic is included only if the organisation that produced it can be named. Where a figure comes from a company measuring its own property, that is stated, because self-reported single-site data is directionally useful and not generalisable. Where a figure is an estimate by a third party rather than a disclosure by the operator, it is labelled as such.

Nothing here is first-party data from me. This page is an index of other people's work, which is exactly what it should be.

Audience and adoption

#	Statistic	Source and period
1	ChatGPT reached 900 million weekly active users	OpenAI, February 2026
2	Google AI Overviews reached roughly 2 billion monthly users	Google, Q1 2026
3	Meta AI passed 1.2 billion monthly active users	Meta, Q4 2025
4	The Gemini app reached 750 million monthly active users	Google, Q4 2025
5	Perplexity reached approximately 45 million monthly active users	Third-party estimate, early 2026
6	Claude reached approximately 30 million monthly active users	Third-party estimate, early 2026
7	On iOS, ChatGPT ranked first, Claude second, and Gemini third by free downloads	SensorTower, early May 2026
8	On iOS top-grossing, ChatGPT ranked first, Claude second, Grok third	SensorTower, early May 2026

Entries 5 and 6 are estimates rather than disclosures, and the gap between Claude's estimated 30 million users and its measured B2B referral output is the most interesting thing on this page.

Referral share and traffic

Statistics 9 through 18 come from Goodie's 2026 AI Search Traffic Report, built on a GA4 panel of anonymised B2B brand sites measured between August 2025 and May 2026. Wave 1 covered 41 brand sites and 2,802,519 AI sessions. It is a B2B panel, so it does not describe consumer categories.

#	Engine	May-Aug 2025	Mar-Apr 2026
9	ChatGPT share of AI referrals	89.1%	62.6%
10	Claude share of AI referrals	1.4%	18.5%
11	Gemini share of AI referrals	2.4%	10.6%
12	Perplexity share of AI referrals	3.1%	7.3%
13	Copilot share of AI referrals	3.2%	4.0%

14. Of sixteen brands with sufficient Claude volume, fourteen moved upward between waves. The report gives a one-tailed binomial p-value of 0.0021. (Goodie, 2026)

15. Claude held 1.29% of measured platform visits and produced 18.0% of B2B referrals, roughly fourteen times its visit share. (Goodie, Jan to Apr 2026)

16. Gemini held 29.0% of platform visits and produced 10.3% of referrals, underperforming its size. (Goodie, Jan to Apr 2026)

17. Perplexity held 1.85% of visits and produced 7.1% of referrals. (Goodie, Jan to Apr 2026)

18. Grok held 3.5% of platform visits and produced effectively no measurable B2B referral output. (Goodie, Jan to Apr 2026)

Engagement and conversion

19. AI-sourced sessions averaged 58.5 seconds of engagement with a 67.8% engagement rate. (Goodie, Feb to May 2026)

20. Google organic averaged 44.2 seconds and a 61.8% engagement rate over the same window, making AI traffic roughly 30% more engaged. (Goodie, Feb to May 2026)

21. Bing organic averaged 47.9 seconds and a 58.4% engagement rate. (Goodie, Feb to May 2026)

22. AI-referred visitors converted at approximately 4.4 times the rate of standard organic traffic. (Semrush, 2026)

23. On Ahrefs' own website, AI search visitors converted at 23 times the rate of organic: 0.5% of traffic produced 12.1% of signups. (Ahrefs, self-reported single-site data, 2026)

Entries 22 and 23 differ by a factor of five because they measure different things. Semrush aggregates across many properties; Ahrefs reports one site with an unusually technical audience. Both support the direction, neither supports a precise multiplier for your business.

Zero-click and AI Overviews

24. Google AI Overviews appeared on about 16% of searches as of November 2025. (Semrush)

25. AI Overviews coverage grew 58% between February 2025 and February 2026. (Semrush)

26. Google's AI Mode shows a zero-click rate of approximately 93%. (Reported by Superlines, 2026)

27. Pages cited by ChatGPT ranked at position 21 or worse in conventional Google results almost 90% of the time. (Semrush, July 2025)

Statistic 27 is the most strategically important entry on this page and the least quoted. It says AI citation is not a reward for ranking. A retrieval system reaching past the first two pages of Google means conventional rank is a weak predictor of citation, which is precisely why passage structure matters more than domain position.

Crawler traffic and blocking

28. AI crawlers accounted for 20.3% of verified bot traffic, with AI-search bots adding a further 6.5%, putting AI-related activity at roughly 26.7% of verified bot traffic. (Cloudflare, May 2026)

29. GPTBot accounted for 11.48% of AI bot HTTP requests. (Cloudflare, May 2026)

30. ClaudeBot accounted for 9.73% of AI bot HTTP requests. (Cloudflare, May 2026)

31. Training-purpose crawling accounted for nearly 80% of AI bot traffic, while user-action and undeclared purposes together stayed under 5%. (Cloudflare, August 2025)

32. In news and publications, GPTBot represented 17.4% of AI crawler activity and ChatGPT-User 14.9%. (Cloudflare, August 2025)

#	Operator	Crawl-to-refer, all sites	Computer and electronics
33	Anthropic	50,000:1	8,800:1
34	OpenAI	887:1	401.7:1
35	Perplexity	118:1	88:1

(Cloudflare, first week of August 2025.)

36. GPTBot appeared in 5.52% of robots.txt DISALLOW rules, the most-blocked AI crawler, ahead of CCBot at 5.08%, ClaudeBot at 4.88%, and Google-Extended at 4.44%. (Q1 2026 analysis of robots.txt across Cloudflare's network)

37. 35.7% of the top 1,000 websites blocked GPTBot. (Originality.AI, 2024)

38. The GPTBot block rate across a sample of 140 million websites was 5.89%. (Ahrefs, 2025)

Statistics 37 and 38 are not in conflict. Large publishers block at far higher rates than the long tail, which is why a single "how many sites block AI crawlers" number is close to meaningless without knowing which sites.

39. Cloudflare began blocking AI crawlers by default for new domains on July 1, 2025. (Cloudflare)

40. On September 15, 2026, Cloudflare sets new defaults across three AI traffic categories, with Training and Agent blocked by default for new domains on ad-displaying pages and Search allowed. (Cloudflare)

Content signals and citation behaviour

41. Adding quotations, statistics, or citations each raised visibility in generative engine answers by roughly 25 to 40 percent, with the three strongest methods achieving 30 to 40 percent relative improvement on a position-adjusted word count metric. Tested across approximately 10,000 queries and nine datasets. (Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024, arXiv 2311.09735)

The important caveat, which almost every guide quoting "40%" omits: the study measured how well a source competes once it is already in a fixed retrieval context. It establishes nothing about whether you get retrieved in the first place, and nothing about durable traffic.

42. Content cited in AI answers was 25.7% fresher than content in traditional organic results. (Ahrefs, 2025)

llms.txt adoption

Three supplementary figures, included because llms.txt is the most-recommended and least-evidenced tactic in the field.

• llms.txt files grew from 4,088 in June 2025 to 36,120 in May 2026 across a monitored set of more than 3 million websites, an 8.8x rise. (Originality.ai)
• 97% of llms.txt files received zero requests in May 2026, across a server-log analysis of 137,000 domains. (Ahrefs)
• Among files that were fetched, SEO audit tools accounted for 21.7% of requests, GPTBot 4.51%, ClaudeBot 0.80%, and AI retrieval bots in aggregate 1.1%. (Ahrefs, May 2026)

The longer version of what this means is in the llms.txt guide.

What was excluded, and why

The following circulate widely and were left out. Each is quoted constantly, and none has a published method behind it.

• "Passages of 40 to 75 words are cited 3.1x more often." Attributed to a 2025 analysis of 10,000 AI citations that has never been published.
• "Pages with tables are cited 4.2x more often than prose." Same origin, same problem. The structural argument for tables is sound without it.
• "FAQ sections get cited 200% more often" and "structured data increases citation probability by 195%." Traced to Medium posts citing other Medium posts.
• "Content over 1,500 words has 180% higher citation probability." Same chain.
• "Only 12% of websites are optimised for AI search engines." No definition of optimised, no sample, no measurement.
• Any figure describing "AI search market share" as a share of total search. The denominators are not comparable across sources and the resulting numbers vary by an order of magnitude.

If you are building a slide from AI search statistics, the exclusion list is more valuable than the inclusion list. Quoting an unsourced multiplier to a CFO who then asks for the study is a bad afternoon.

For what to do with these numbers, see what actually changes between GEO and SEO and the extraction patterns.

Frequently Asked Questions

What percentage of AI referral traffic comes from ChatGPT?

62.6% of B2B AI referrals in Goodie's GA4 brand panel for March to April 2026, down from 89.1% in the May to August 2025 wave. Claude was second at 18.5%. The figure covers B2B sites only and excludes Google AI Overviews, where the answer usually ends the session without a referral.

Do AI search visitors convert better than organic?

The available evidence says yes, with wide variance in the size of the effect. Semrush measured roughly 4.4x standard organic across many properties. Ahrefs reported 23x on its own site, where 0.5% of traffic drove 12.1% of signups. Use the direction, not the multiplier.

What did the Princeton GEO study actually prove?

That adding quotations, statistics, and citations each raised visibility by roughly 25 to 40 percent on a position-adjusted word count metric, tested over about 10,000 queries across nine datasets. The widely quoted 40 percent is a maximum, not an average, and the effect was measured on sources already present in a fixed retrieval context, so it says nothing about whether you get retrieved at all.

How many websites block AI crawlers?

It depends entirely on which websites. Originality.AI found 35.7% of the top 1,000 sites blocking GPTBot in 2024, while Ahrefs measured a 5.89% block rate across 140 million sites in 2025. Large publishers block at far higher rates than the long tail, so any single headline number is misleading without the sample.

Why is AI search traffic underreported in analytics?

Users copy URLs out of answers or type the domain later, which registers as direct traffic, and native mobile apps frequently strip the referrer. Separately, Google bundles its AI surfaces into organic attribution, so an AI Overviews visit is indistinguishable from a blue-link visit in Search Console.

Does ranking on Google predict AI citation?

Weakly. Semrush found in July 2025 that pages cited by ChatGPT ranked at position 21 or worse in conventional results almost 90% of the time. Retrieval systems reach well past the first page, which is why passage-level structure matters more than domain position.

Which AI search statistics should I not quote?

Any specific citation multiplier attached to passage length, table presence, FAQ sections, or word count. Those figures trace back to unpublished analyses and Medium posts citing each other. The same applies to any claim about AI's share of total search volume, since the denominators are not comparable across sources.