For two years the AI-referral conversation has run on anecdotes: someone's client saw a spike from ChatGPT, someone else swears Perplexity converts. Now there is a dataset worth arguing over. Previsible's David Bell, writing at Search Engine Land, analysed 6.77 million LLM-referred sessions across 166 GA4 properties, spanning November 2024 to May 2026 and covering SaaS, e-commerce, finance, legal, health, education and publishing. The same 166 sites are present for the whole 19-month window, which makes the trends unusually clean. Three findings matter for anyone who owns a traffic number.

Finding one: it is growing fast, and it is basically all ChatGPT

Monthly LLM referral sessions across the panel grew 9.9x over the window, from 65,249 in November 2024 to 644,478 in May 2026. And that growth is not spread across the field: ChatGPT accounts for 92.4% of trackable LLM referral traffic, up 12.8x over the period with, in Bell's words, no sign of slowing.

The rest of the podium is small and getting smaller. Gemini holds a steady ~3.2%. Perplexity is down 61% from its March 2025 peak to roughly 2.7%. Copilot has collapsed 96% from its August 2025 high to a rounding error. The one genuine mover is Claude: from 133 sessions in November 2024 to 8,528 in May 2026, a 64x rise that saw it overtake Perplexity in March 2026, driven largely by enterprise adoption. If your audience is technical or professional-services, Claude is the early-positioning opportunity in this data.

Finding two: this traffic can halve overnight

The growth curve is not a smooth line. Between August and November 2025, panel-wide sessions dropped from 396,278 to 213,345, a near-50% fall, before recovering and pushing on to new highs. The cause was product changes at a single vendor, not anything the 166 sites did.

When one platform is 92% of a channel, one product decision at that platform is 92% of your risk in it.

That is the strategic read: AI referrals are worth capturing, but they are a concentrated, volatile dependency. Treat them the way a sensible finance team treats a customer who is 92% of revenue: welcome the money, and do not build the forecast on it.

Finding three: a quarter of it lands on your internal search

This is the finding most teams will have missed, and the most actionable. Roughly 25% of AI-referred traffic lands on internal search results pages, rising to 28.8% for ChatGPT specifically. The models frequently know which domain to send someone to, but not which page, so they hand the visitor to your own site search and let it finish the job.

The pattern varies sharply by vertical, and each variant is a to-do list:

  • SaaS: 34.6% of LLM traffic lands on internal search pages. If your site search is an afterthought, a third of this channel arrives at your weakest page.
  • E-commerce: 43% lands on product pages, so the models are deep-linking. Product page quality and structured data are your AI landing experience.
  • Education: 52% lands directly on course pages; health: 42.1% on About pages, which suggests the models route users to trust signals before treatment content.
  • Publishers: 54% lands on news pages, but penetration was just 0.11% against 120+ million organic sessions. For content businesses this channel is still a drop in the ocean.

Bell's data also shows a split in platform behaviour worth knowing: ChatGPT and Gemini act on domain trust and often defer to your internal search, while Perplexity and Claude tend to select specific content pages. Different engines, different landing experiences to audit.

The honest caveats

Two things keep this study from being a mandate to reorganise your marketing. First, scale: even after 9.9x growth, AI referrals remain a small fraction of organic for most sectors, publishing's 0.11% being the starkest example. Second, the study measures sessions, not outcomes; conversion rates by LLM platform remain unmeasured. There is decent third-party evidence elsewhere that AI-referred visitors are high-intent, but this dataset does not prove it, and neither should your board deck.

What to do this month

  • Instrument the channel. Segment LLM referrers (chatgpt.com, gemini.google.com, perplexity.ai, claude.ai) in GA4 so you have your own baseline, your own volatility picture, and eventually your own conversion answer to the question the study could not.
  • Audit your internal search like a landing page. Run the queries an AI-referred visitor would arrive with. If site search returns junk, has no analytics, or dead-ends on zero results, you are burning a quarter of the channel on arrival.
  • Check where the models actually send people. Pull your top LLM landing pages. If they are About pages or thin search results rather than money pages, tighten the pages the models trust and interlink them to the ones that convert.
  • Watch Claude if you sell to technical buyers. A 64x riser with an enterprise skew is exactly where being cited early is cheap and defensible.

Where we come in

Our AI Visibility Audit answers the questions this study raises for your specific site: which assistants send you traffic today, whether you or a competitor is the default answer in your category, and where AI-referred visitors land, including whether your internal search helps or haemorrhages them. The study gives you the market picture. The audit tells you your position in it, and what to fix first.