ChatGPT drives about 92.4% of standalone-LLM referral traffic, according to Previsible's 2026 State of AI Discovery report, published 6 July 2026, but that figure does not mean it is the only engine worth optimising for. Previsible's own report notes that AI discovery happening inside Google, including AI Overviews, almost certainly sends more traffic than all standalone LLM platforms combined. So 92% is ChatGPT's share of one large slice, not of all AI discovery.
What this data is, and isn't. Previsible analysed 6.77 million LLM-driven sessions across 166 GA4 properties in nine industries, from November 2024 to May 2026. It measures referral clicks that arrive tagged from a standalone chat product, not citations, not crawler hits, not business spend, and not the huge volume of zero-click AI answers that never send anyone anywhere. It is one analytics dataset, directional and skewed to sites that tag AI referrals well.
What did Previsible actually measure?
Standalone-LLM referral sessions. Visits where a chat assistant sent a click to a website. Aggregated across GA4:
| Metric (Previsible 2026 State of AI Discovery) | Figure |
|---|---|
| Sessions analysed | 6.77 million |
| GA4 properties / industries | 166 / 9 |
| Date range | Nov 2024 - May 2026 |
| Monthly session growth | ~9.9x, to 644,478 (May 2026 peak) |
| ChatGPT share of standalone-LLM referrals | ~92.4% |
| Claude growth | ~64x, to 8,528/mo; overtook Perplexity Mar 2026 |
| Nov 2025 dip | ChatGPT referrals fell 448,412 → 213,345 in one month |
Source: Previsible 2026 State of AI Discovery report, reported 6 July 2026. One scope caveat worth keeping: analytics tools misattribute a lot of AI referrals as "direct," so even this fast-growing number is an undercount. The mechanics are in why AI traffic shows up as direct.
Does 92% mean I can ignore other engines?
No, and the same report gives three reasons.
First, Google is the bigger discovery layer. Previsible states plainly that AI discovery inside Google's results likely exceeds all standalone LLMs combined. This standalone figure simply doesn't include it, which is why the corpus treats Google's AI surfaces as a first-class target (see Google AI Mode usage data and AI Mode vs AI Overviews citations).
Second, referral share is one metric among several, and each has a different leader. "Which AI is winning" depends entirely on what you count:
| Metric | What it counts | Source (mid-2026) | Who leads |
|---|---|---|---|
| Standalone-LLM referral traffic | Clicks sent to sites | Previsible, Jul 2026 | ChatGPT ~92.4% |
| AI crawler traffic | Bot requests hitting sites | Cloudflare, mid-2026 | Googlebot ~24.9%, ClaudeBot ~20% |
| Business AI adoption | Paid business spend | Ramp, Apr 2026 | Anthropic 34.4% > OpenAI 32.3% |
| Sources cited per answer | Citation behaviour | Semrush, 2026 | Varies (ChatGPT ~15, Gemini ~3) |
These do not contradict each other. They measure different things. ChatGPT can lead referral clicks while ClaudeBot leads purpose-built crawling (see AI crawler market share) and Anthropic leads business spend (see Anthropic overtakes OpenAI in business adoption). Reading one as "the" AI leaderboard is the misread to avoid.
Third, referral volume is volatile. Previsible recorded a roughly 50% single-month drop in November 2025, driven almost entirely by ChatGPT. A number that can halve in a month is a shaky foundation for a single-engine strategy.
ChatGPT wins the referral click today. But referral clicks, crawler hits, business spend, and citations each crown a different engine, so "optimise for the leader" is a question, not an answer. Measure all four for the engines your buyers use.
Why is Claude overtaking Perplexity in referrals?
Because its rise in referrals tracks its rise in adoption. Previsible reported Claude grew about 64x over the period, reaching 8,528 monthly sessions and passing Perplexity in March 2026. Flat through 2025, then accelerating as agentic and enterprise tools landed. That is the same direction as Ramp's spending data without restating it: attention is broadening beyond a single assistant. Claude is still a small referral share, so treat it as a fast-rising signal to add to measurement, not a new leader to chase.
How much AI referral traffic am I probably missing?
More than your dashboard shows. Because assistants often send visits that land as "direct," the true figure is higher than tagged referrals suggest. A form of dark traffic and, for pages read-but-not-clicked, the dark library effect. The fix is measurement built for it: a referral view that separates AI sources (GA4 AI-referral dashboard) and, more importantly, presence and citation tracking across every engine, since a zero-click answer can build your brand without ever sending a session.
The honest read of Previsible's data is that AI referral traffic is real, growing fast, ChatGPT-led, and only part of the picture. Knowing your share of citations and recommendations across ChatGPT, Google's AI surfaces, Claude, and Perplexity, not just the clicks one of them sends. Is what Buffy Intel tracks daily, so you optimise for the whole board rather than one square. Questions: [email protected].