AI Assistant Referrals Across 31 Websites: The 2026 Numbers.

Across the 31 websites Theory Road owns and operates, AI assistants sent 10,194 sessions in the 90 days from July 18 to October 16, 2026. That is 10.1% of all 101,140 sessions and 23.8% of what Google organic search sent in the same window. One AI assistant visit arrived for every 4.2 Google visits. ChatGPT sent 93% of it.

This is our second look at AI referrals. The first covered one site and its lead data: 77 ChatGPT sessions, 4 leads, 5.2% against 2.4% from Google. This one is the whole portfolio. Client sites are excluded on purpose, each property is described only by type, and the numbers come straight from GA4.

The headline numbers.

Five channels account for nearly all traffic. Direct is inflated by bots on several of these sites, so it stays out of every ratio here. The comparison that matters is AI Assistant against Organic Search: both answer the same intent, someone asked a question and a machine pointed them at a page.

The 4.2 to 1 ratio is the number we track. Session share moves with paid spend and bot traffic; AI sessions divided by organic sessions does not. Pull both channels for your own site over the same 90 days and compute that ratio first.

Who sends it.

ChatGPT sent 9,466 of the 10,194 AI sessions, or 92.9%. Gemini is a distant second at 289, then Claude, Perplexity and Copilot between 119 and 143 each. The five others combined: 677 sessions, under 7% of the channel.

Two things follow. No separate playbook per assistant is needed; ChatGPT is the only one with enough volume to measure, and the pages it cites are the pages the others cite. And Gemini is the one to watch, because it sits inside Google's ecosystem and could change the mix fast. Optimize for the question, not the chatbot.

Which kinds of sites get AI traffic.

AI referrals are not spread evenly. 23 of the 31 sites received at least one AI session. 14 have an AI share of 5% or more, 8 are at 10% or more, and the median site sits at 4.5%. The 8 sites with zero AI referrals all had fewer than 250 sessions in the window, so they are new or tiny.

A small directory and data site is at 48% AI share (43 of 89 sessions, against 2 from Google). A review site is at 42%, with 799 AI sessions against 341 organic. Another directory and data site is at 27% (343 against 219) and a magazine is at 21% (139 against 114). On all four, AI assistants sent more visitors than Google. The largest property, a review and comparison site with 44,332 sessions, sits at 11.6%: 5,128 AI against 18,735 organic.

The 42% and 48% sites have four things in common. The topic is narrow. There is one page per question, so the URL that answers 'is X good for Y' answers only that. Every comparison lives in a real table with the same columns on every page. And the first paragraph gives the plain answer before any explanation. That is the shape a language model wants to quote. Local business sites are the opposite case: people ask a chatbot for a place to go and get a map, not a link. If that is your business, set expectations from the last row.

The trend since April, with a caveat.

GA4 did not have an AI Assistant channel until this spring. Google's What's new in Google Analytics page carries the entry dated May 13, 2026: sessions from recognized assistants now get the medium 'ai-assistant' and land in a new AI Assistant default channel group. Before that they were classified as Referral or Unassigned. So April is zero by definition, May is partial, and the clean comparison starts in July.

Two things are moving at once. The early months show the channel being adopted inside GA4. From July on, the AI count rose every full month: 2,842, then 3,013, then 3,793. The ratio fell because organic grew faster, and organic grew faster because we published several hundred pages in August and September. A denominator effect, not a decline in AI visits.

October deserves a plain reading too. The first 16 days ran at roughly 100 AI sessions a day against about 126 a day in September. Sixteen days is not enough to call a plateau; the next edition will say. For your own reporting: never read AI share without the AI count next to it.

What it means for a business site.

AI traffic is small but real. On a content-heavy portfolio it is one session in ten. On a service business site it will be less: our two agency and service properties sit at 7.4% on only 74 sessions, and the local business sites are near zero. If you see 2% to 5% today, you are normal. If you see zero, confirm your GA4 property shows the new channel first.

It clusters on specific pages. On every property in this set, a handful of pages carry most of the AI sessions, and they are the pages that answer one question completely. And it is cheap to earn relative to Google, because the pages that win are the same pages. Nothing in this dataset came from a separate AI budget or team. AI referrals are a byproduct of doing content and technical SEO properly, not a new channel you buy into. Treat the AI Assistant row as a quality signal and act on the pages that earn it.

What this index cannot tell you is whether AI visitors convert. Most of these content sites have no key events configured, so we make no conversion claim for the portfolio. The only conversion data we have is the single-site piece linked above: 4 leads from 77 ChatGPT sessions. One site, small numbers. Read it as a hint, not a benchmark, and set up your own key events.

What we do on every property to earn it.

The same checklist runs on all 31 sites. Most of it is what good SEO already looked like, plus two additions that cost an hour. The full method is in our guide to getting found by AI search. The short version:

How to measure it yourself.

Carry the caveats with whatever you measure. GA4 attributes by referrer, so in-app browsing that strips the referrer lands in Direct and is undercounted. Direct is bot inflated on content sites, so we never use it as a denominator. And this is one operator's portfolio, heavy on review and comparison content. The method is portable; the percentages are ours.

What we are doing next.

This index updates quarterly. The next edition covers October 17, 2026 to January 16, 2027, with the same properties and the same tables so the numbers are comparable. If the October slowdown is real, it will show. If Gemini or Claude takes share from ChatGPT, it will show.

The bigger change is conversion tracking. Between now and January we are configuring key events on the content sites: outbound clicks, form submits, and calls. The next edition can then report leads per channel and test whether the 5.2% lead rate holds across a portfolio. We run this stack for the businesses we work with. If you want it on your site, with the reporting to prove it, Hire Theory Road. Questions about the data go to Travis directly.

Questions we get about this data.