AI traffic in GA4: what the new "AI Assistants" channel shows - and what it misses

In mid-May 2026 a new channel appeared in the acquisition reports of many GA4 properties: "AI Assistants". No banner, no setup - Google switched the assignment on server-side on 13 May 2026. For the first time, GA4 answers on its own how many visitors arrive from ChatGPT, Gemini and co.

The short answer now sits in the report. The honest answer is longer: the channel captures only part of your AI traffic, it starts at zero, and of all things, Google's own AI answers do not count towards it. This post covers both - what the channel can do and where you have to measure yourself.

What Google introduced on 13 May 2026

The changelog entry is titled, plainly, "New AI Assistant traffic measurement" (What's new in Google Analytics, 13 May 2026). Behind it sit three automatic assignments:

The channel rule itself is remarkably plain: the channel applies when the medium exactly matches ai-assistant - that is how the default channel group definition puts it (as of 17 Aug 2026). Which referrers Google maps onto that medium is decided by an internal list. It is not published.

Two details are missing from almost every post on the topic. First, Google itself is inconsistent about the name: the changelog says "AI Assistant", the channel definition says "AI Assistants". Second, the rollout was staggered - announced on 13 May, broadly distributed across properties only around 7 June 2026. If your property showed nothing in May, that was not a bug. That was the rollout.

Where to find your AI traffic

Fastest route: Acquisition → Traffic acquisition. "AI Assistants" appears as a row in the "Session default channel group" dimension as soon as recognised visits come in. In explorations the channel is available through the same dimension; in the first-user scope it arrived later.

Since 11 June 2026 there is also a second lens: the changelog update "New Source Group field" bundles sources into groups - and there, Google lists "ChatGPT (OpenAI)" and Perplexity as groups of their own. Keep that difference in mind, it matters in a moment: Perplexity exists in the Source Group field, but not in the AI channel.

Which assistants Google recognises - and which it does not

Both Google lists are explicitly phrased as examples ("sources like ..."); a complete list does not exist. The drift is striking: the 13 May announcement named ChatGPT, Gemini and Claude - today's definition page names ChatGPT, Gemini, Deepseek, Copilot and Grok. Claude has vanished from the examples, Perplexity was never in. A hands-on test by Clickport (updated 12 June 2026) confirms it: a perplexity.ai referrer is classified as Referral, not as an AI assistant. If you miss an assistant, you can report it to Google - inclusion requests go to googleanalytics-channelgrouprequests@google.com (documented by Search Engine Journal, July 2026).

And then there is the fine print almost nobody quotes: according to the channel definition, clicks from Google's AI Overviews and AI Mode count towards Organic Search - not towards the AI Assistants channel. If you read the channel row as "total AI influence", you are missing exactly the AI traffic that comes through Google Search itself.

Why the number is a floor

Channel detection hangs on the referrer header. And that header is missing for a substantial share of AI visits: clicks from native apps (the ChatGPT app, Gemini on iOS), from in-app browsers and from pages with a strict referrer policy arrive with no sender. GA4 then files them under Direct - invisible to the new channel.

Several providers have measured the size of that gap on their own data - and the results cluster tightly. Loamly found across 446,000 examined visits that 70.6 % of AI traffic arrived without a referrer (as of February 2026). Attrifast measures in a 200-site cohort that GA4 files roughly 70 % of ChatGPT sessions under Direct (May 2026) - the ChatGPT web interface passes the referrer on about 28 % of clicks, the desktop app on 6 %. The widely quoted "20-40 %" range, by contrast, has no traceable original study; the published measurements sit well above it. Small studies confirm the pattern: Wheelhouse DMG found 56 Gemini visits from iOS in server logs - 5 of them arrived in GA4 with a referrer (cited via Clickport, June 2026).

Two structural blind spots come on top, and no channel will ever close them: users who copy a URL out of the AI answer instead of clicking never produce a referrer. And ad blockers plus privacy tools strip the header even where it would be sent.

There is partial compensation: ChatGPT appends utm_source=chatgpt.com to links it outputs - but not everywhere and not always. Citations are usually tagged, links in search-result lists partly are not, and the behaviour has changed several times since 2024.

Which leads to the next trap: even the ChatGPT traffic that does arrive does not land in the AI channel as one block. Depending on whether the referrer is recognised, only the UTM parameter survives, or both are missing, the same source spreads across up to three channels - AI Assistants, Referral and Direct, plus Unassigned when utm_source arrives without a valid medium (Search Engine Journal, 15 July 2026). For numbers you can trust, filter across all channels on the source chatgpt.com instead of reading the channel row alone. For every other assistant the advice is: watch your Direct share. If it grows in step with your AI visibility, AI traffic is probably hiding there - the same logic you use to break down unassigned traffic and (not set) sources.

Why measuring this is worth the effort

Not because of the volume - that is still small. Ahrefs put the measurable AI share across 3,000 websites at 0.17 % of visitors (February 2025), and Similarweb's 1.1 billion AI referrals in June 2025 (+357 % year over year) are still a fraction of the 191 billion coming from Google Search.

It is worth it because of the quality. At Ahrefs itself, AI search brought 0.5 % of visitors but 12.1 % of signups - a 23x higher conversion rate than traditional organic search (June 2025). Semrush values AI visitors, measured by conversion rates, at 4.4 times an organic search visitor (study, June 2025). Adobe measured a 42 % better conversion rate in March 2026 - after AI traffic had still converted worse than other channels a year earlier. The bitterest punchline comes from Loamly: AI traffic disguised as Direct converted at 10.2 % in their data, against 2.5 % for the rest. The most valuable visitor cohort is hiding in exactly the channel nobody looks at.

13 May 2026 is your baseline zero

GA4 does not reclassify historical data. Sessions before 13 May 2026 stay in Referral or Direct permanently - the channel only counts forward. Google says nothing about this in the changelog; the non-retroactivity is documented in practice and matches how default channel groups behave.

Three concrete moves follow from that:

1. Set an annotation on 13 May 2026. Any before/after comparison without that marker is misleading.
2. Expect an apparent Referral drop. Traffic that used to run as Referral moves into the new channel from May onwards. A year-over-year comparison of the Referral channel is worthless from now on without that lens.
3. Build the historical baseline with a custom channel group. According to Google's documentation, custom channel groups apply to reports retroactively - they are the only way to make AI traffic visible before 13 May 2026.

Do you still need the manual channel group?

Yes - for the history and for the gaps. As a base for your own rule and for BigQuery analyses: the common referrer domains of the assistants (as of July 2026, documented by DarwinApps):

A custom channel group with a "source matches regex" rule covers history and gaps in one:

chat\.openai\.com|chatgpt\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|grok\.|x\.ai|deepseek\.com|meta\.ai|you\.com

Three pitfalls from practice:

• Rule position: GA4 assigns traffic to the first matching channel ("first match"). The AI rule therefore belongs above Referral, otherwise Referral keeps swallowing the Perplexity visits.
• Extend the medium condition: since 13 May, GA4 rewrites the medium of recognised assistants to ai-assistant. An older rule that checked medium = referral plus source chatgpt.com no longer matches new ChatGPT traffic. No Google doc states this consequence; it follows directly from the documented mechanics.
• Mind the limit: standard properties allow only 2 custom channel groups (360: 5). Do not burn a slot on a copy of the default group.

What to audit in your own property now

1. Check the start date of the AI rows. Filter traffic acquisition on "AI Assistants" and check when data starts flowing - mid-May to early June is normal; a later start points to little recognised AI traffic.
2. Hold your Direct share against thresholds. The GA4 Auditor rates the Direct share with <20 % green, 20-35 % amber, >35 % red - the reasoning is in the help page on attribution & acquisition quality. A growing Direct block is the most likely home of invisible AI traffic.
3. Cross-check your referral exclusions. If an assistant domain ended up in "List unwanted referrals" by accident, its traffic turns into Direct - and the channel stays empty.
4. Verify at event level. In the BigQuery export, the channel assignment sits per event in the field

session_traffic_source_last_click.cross_channel_campaign.default_channel_group

More importantly, the export preserves source and referrer for every event. That lets you rebuild the AI classification retroactively - exactly what the GA4 UI cannot do for the default channel group:

SELECT
  session_traffic_source_last_click.manual_campaign.source AS source,
  COUNT(DISTINCT CONCAT(user_pseudo_id, (
    SELECT value.int_value FROM UNNEST(event_params)
    WHERE key = 'ga_session_id'))) AS sessions
FROM `project.analytics_XXXXXX.events_*`
WHERE REGEXP_CONTAINS(
  session_traffic_source_last_click.manual_campaign.source,
  r'chatgpt|openai|perplexity|claude\.ai|gemini|copilot|deepseek|grok')
GROUP BY source
ORDER BY sessions DESC

The GA4 Auditor reads the channel distribution from both worlds (Data API and the BigQuery export) and flags elevated Direct and Unassigned shares automatically. Check your property now.

Related topics

• Unassigned traffic in GA4: finding and fixing the causes
• "(not set)" in GA4 - where auditing really starts
• Help page Attribution & acquisition quality
• GA4 audit checklist

#attribution#ai-traffic#channels#ga4