A retainer this practice closed in 2026 started with a prospect who already knew the work before the first call. He had not found the site through Google. He had asked an AI assistant to name someone who could own a B2B SaaS marketing site, read the answer, and reached out. In the analytics, that visit was filed under Direct, indistinguishable from someone typing the URL. The highest-intent lead of the quarter was invisible at the exact moment it mattered.
AI Assistants Became a Referral Channel While GA4 Looked Away
AI assistants are now a measurable referral channel, and it grew faster than almost anything in marketing analytics. In June 2025, AI platforms sent 1.13 billion referral visits to the top 1,000 websites, up 357% year over year, with ChatGPT alone driving more than 80% (Similarweb, 2025)[1]. The channel exists, it is compounding, and most marketing dashboards have no row for it.
Two honesty checks keep this from becoming hype. First, the absolute volume is still small. Across 2025 the channel grew 66%, from 462 million to 767 million monthly visits, but that is under 0.15% of total web traffic (Semrush, 2025)[2]. Second, the more important number is reach, not referrals. In 2025, Google reported its AI Overviews at 2 billion monthly users and AI Mode at 100 million[3]. Far more buyers are reading AI answers about your category than are clicking through from them.
That reach-without-clicks gap is the real story. The assistant answers the question inside the chat, the buyer forms an opinion, and no visit gets logged. For a marketing team, the channel is not just undercounted. It is partly invisible by design, because the product is built to resolve the query without sending a click.
Why GA4 Files AI Traffic as “Direct”
GA4 misfiles AI traffic for a specific technical reason: the apps strip the referrer on the way out. When someone taps a link inside the ChatGPT or Perplexity mobile app, the in-app browser frequently drops the Referer header. With no referrer, GA4 has nothing to attribute the session to, so it lands in Direct next to bookmarks and typed URLs. The visit is real. Its origin is erased before GA4 ever sees it.
The undercount compounds with the zero-click reality of modern search. In a 2025 Pew Research study of nearly 69,000 searches, users clicked a result on just 8% of searches that showed an AI summary, versus 15% without one, and clicked a link inside the summary on only 1% of visits[4]. Zero-click was already the norm before AI: in 2024, 58.5% of US Google searches ended without a click of any kind (SparkToro, 2024)[5]. AI answers widen a gap that was already there.
This is the same failure mode as the Analytics Trust Gap, pointed at a new channel. The dashboard does not show you what it missed. It shows a Direct number that quietly absorbs your AI-referred buyers, and the team reads that number as people who typed the URL. Attribution that is wrong by an unknown amount is worse than attribution the team knows is wrong, because nobody adjusts for it.
The Buyers You’re Missing Are the Ones Closest to Buying
The traffic GA4 is burying is disproportionately high-intent, which is what makes the blind spot expensive. In a 2025 Gartner survey of B2B buyers, 45% said they used generative AI to research vendors and products, and 69% preferred to validate AI-generated insights with a sales rep before deciding[6]. The assistant is now part of the buying committee’s research step, and it hands off warm.
Behavioral data points the same way. Adobe Analytics found that visitors arriving from generative-AI sources browse 12% more pages per visit and bounce 23% less than other traffic (Adobe, 2025)[7]. That is an engagement and research-depth signal, not a promise of higher last-click conversion. The right read is that these visitors arrive further along, having already used the assistant to shortlist. They behave like people checking a recommendation, not people starting a search.
The prospect from the opening was exactly this pattern. He arrived pre-sold because the assistant had done the shortlisting, and the only reason the source was knowable at all was that he said so on the call. Multiply that by a quarter of demo requests and the cost of the blind spot is clear. You cannot double down on a channel your reporting cannot name, and you cannot defend its budget when finance asks where the pipeline came from.
How to Make AI Referrals Visible in GA4
You cannot make AI traffic perfectly visible, but you can move it from invisible to directional with three steps. The point is a defensible read, not a precise count, and admitting that boundary is what separates an honest setup from the guides that overpromise.
First, build a custom channel group. In GA4, under Admin and Channel groups, add a channel matched with a regex on the source field covering openai.com, chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, and copilot.microsoft.com. That reclassifies every AI session that did arrive with a referrer into one named bucket, instead of scattering it across Organic and Referral. It will not touch the no-referrer sessions already sitting in Direct, but it captures the share that came through intact.
Second, watch the proxies. Segment Direct traffic to landing pages that are not your homepage, because a buyer who types your URL lands on the homepage, while a buyer following an AI link often lands deep on a specific page. Track branded-search volume in Search Console as a parallel signal: when assistants mention you, branded searches tend to rise even when clicks do not. Neither is exact. Together they triangulate.
Third, ask. Add a single open field to demo and contact forms (how did you first hear about us) and read the answers monthly. Self-reported attribution is messy, but it is the only source that captures the zero-click buyers who never generated a session at all. Pair this with the 30-minute GA4 audit so the channel-group work sits on a collection layer you have already checked. And hold the whole picture loosely, because operating well on data you know is incomplete is the entire argument behind what data-driven actually means when your data is wrong.
Sources
- Similarweb (via TechCrunch), AI referrals to top websites up 357% year over year – Similarweb clickstream, top 1,000 domains, June 2024 vs June 2025; 1.13 billion AI-referral visits in June 2025; ChatGPT drove 80%+ ↩
- Semrush, How AI is reshaping traffic channels – Billions of visits across 50,000+ sites and 17 industries, Jan-Dec 2025; AI traffic grew 66% (462M to 767M monthly visits) but remains under 0.15% of total web visits ↩
- Google/Alphabet Q2 2025 earnings (via TechCrunch), AI Overviews reach 2B monthly users – Reported July 23 2025; AI Overviews at 2 billion monthly users (up from 1.5B in May 2025); AI Mode at 100M monthly active users in the US and India ↩
- Pew Research Center, Google users are less likely to click when an AI summary appears – 68,879 searches from 900 US adults tracked March 2025; users clicked a result on 8% of searches with an AI summary vs 15% without; summary links clicked in 1% of visits ↩
- SparkToro + Datos, 2024 Zero-Click Search Study – Datos clickstream panel, tens of millions of panelists, Sept 2022-May 2024; 58.5% of US Google searches ended in zero clicks (no click of any kind) ↩
- Gartner, 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights – Survey of 645 B2B buyers fielded Aug-Sept 2025; 45% used generative AI to research vendors and products; 69% prefer to validate AI insights with a sales rep ↩
- Adobe Analytics, Traffic from generative AI sources jumps 1,200% – Adobe Analytics across 1T+ visits plus a 5,000-consumer survey, data through Feb 2025; generative-AI visitors browse 12% more pages and bounce 23% less than non-AI traffic ↩
Seeing these patterns at your company?
Book a free WebOps Diagnostic. I'll review your site before the call and share specific observations.
Book a Free Call →Frequently Asked Questions
Because the ChatGPT mobile apps open links inside an in-app browser (a WebView) that often drops the Referer header on the outbound tap. With no referrer, GA4 has nothing to attribute the session to, so it files it under Direct alongside bookmarks and typed URLs. Perplexity, Gemini, and Copilot apps behave similarly. The traffic is real and often high-intent; GA4 just cannot see where it came from without a custom rule.
In GA4, go to Admin, then Channel groups, and create a custom group. Add a channel matched on a regex against the source field covering AI referrer domains: openai.com, chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, and copilot.microsoft.com. This reclassifies the referrer-bearing AI sessions out of Organic and Referral into a named bucket. It cannot recover the no-referrer sessions that already landed in Direct, but it captures everything that arrived with a referrer intact.
In June 2025, AI platforms sent 1.13 billion referral visits to the top 1,000 sites, up 357% year over year, with ChatGPT driving over 80% (Similarweb). Across 2025 the channel grew 66% (Semrush). It is still under 0.15% of total web visits, so the volume is small, but for B2B SaaS the intent quality is high because the visitor arrives after the assistant has already shortlisted vendors.
The honest answer is that it engages more, and conversion depends on the cohort. Adobe Analytics found generative-AI visitors browse 12% more pages and bounce 23% less than other traffic. That signals higher research intent, not a guaranteed higher conversion rate. Treat AI-referred visitors as deeper in the consideration stage, and judge them on assisted-conversion and pipeline contribution rather than last-click conversion alone.
No, and that is the part most guides skip. A custom channel group recovers the sessions that arrive with a referrer, but the no-referrer share is unrecoverable inside GA4, and zero-click answers mean many buyers read about you in an AI summary and never click at all. The realistic goal is a directional read using channel groups, Direct and branded-search trends, and demo-form self-reporting, not a precise count.
