GA4’s AI Assistant Channel vs Your Own Channel Group: Which AI-Referral Number Should You Trust?
Short answer: neither in isolation. GA4’s native channel is called AI Assistant (singular), landed on 13 May 2026, and classifies purely by referrer. Google’s own two help pages name different source lists and never publish the full one. A custom channel group applies retroactively; the native channel is not documented to. Reconcile the two, report the wider figure, and treat every number as a floor.
If you built a custom channel group for AI traffic, you now have two numbers on one dashboard measuring roughly the same thing, and they disagree. This post is about which one goes in the board deck. It is the companion to our setup guide on measuring AI referral traffic in GA4: that post builds the layers, this one covers what to do when they argue.
What Google actually documents, as at 26 August 2026
The name is singular. Google’s [GA4] Default channel group help page says: “AI Assistant is the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok. It excludes Google’s AI Overviews and AI Mode.” If you are writing a filter or a BigQuery query against the channel string, “AI Assistants” will match nothing.
The mechanism is referrer-only. The same page gives the condition as: “The medium exactly matches “ai-assistant”. The medium is set to “ai-assistant” and the campaign is set to “(ai-assistant)” if the referrer matches a list of AI Assistants.” Note the phrasing: a list, which Google does not publish.
Two Google pages give two different lists. Both were live and checked on 26 August 2026.
| Google page | AI sources it names | Notable |
|---|---|---|
| What’s new in Analytics, 13 May 2026 | ChatGPT, Gemini, Claude | Names Claude; omits Deepseek, Copilot, Grok |
| [GA4] Default channel group | ChatGPT, Gemini, Deepseek, Copilot, Grok | Omits Claude; neither page names Perplexity |
Both pages hedge with “sources like” or “a recognized AI Assistant”, suggesting the list is maintained server-side and changes without a changelog. The consequence: you cannot audit what the native channel counts, only observe what it produced.
The backfill question, where most reconciliations break
Google’s [GA4] Custom channel groups page states plainly: “Custom channel groups can be applied to your reports retroactively.” Build one today and your AI channel populates backwards across existing source and medium data. Standard properties allow two additional custom groups, with up to 50 channels each.
The native channel carries no equivalent statement. Neither the release note nor the default channel group page mentions backfill or reprocessing, and we will not claim Google definitely does not backfill, because Google has not documented either way. The rule holds regardless: do not run a year-on-year comparison on the native channel across 13 May 2026 and call it growth. You would be comparing a period where the classifier existed against one where it did not.
Why the two numbers diverge
Assuming both are configured correctly, four mechanisms explain the gap.
1. Source-list mismatch
Your regex names hosts you chose; Google’s list is unpublished. If your pattern covers perplexity.ai and claude.ai, your custom group will exceed the native channel. Usually the largest component of the gap, and the one you can explain.
2. Channel ordering
GA4 evaluates channel conditions top to bottom and stops at the first match. A custom AI channel below Referral reports zero forever, with no error. If your custom number is lower than the native one, check ordering first; the fix is in the setup guide.
3. Referrer stripping
Both methods need a referrer, and often none arrives. Assistant apps, in-app webviews and copy-pasted links produce sessions with no referrer and no UTM parameters, which GA4 classifies as Direct. These are lost identically to both methods, so this inflates Direct rather than widening the gap.
4. Google’s own AI surfaces sit elsewhere by design
The documentation is explicit that AI Assistant “excludes Google’s AI Overviews and AI Mode”, and that Organic Search includes “Google’s AI Overviews and AI Mode”. A click from an AI Overview is Organic Search. If your regex included google.com hosts to catch AI Mode, the numbers diverge for reasons unrelated to assistants.
The two numbers side by side
| Method | What it sees | Blind spots | Historical backfill | Effort |
|---|---|---|---|---|
| GA4 native AI Assistant channel | Sessions whose referrer matches Google’s unpublished list, with engagement and conversion metrics | Referrer-less visits (become Direct); assistants off the list; AI Overviews and AI Mode by design; list not inspectable | Not documented either way; treat pre-13 May 2026 as unclassified | None, on by default |
| Custom channel group (source regex) | Any referrer host you name, as a first-class channel with conversions | Referrer-less visits; assistants not yet added; silently zero if ordered below Referral | Yes, documented as “can be applied to your reports retroactively” | Low to build, ongoing to maintain |
Two further layers sit outside this particular argument, and we have covered both elsewhere rather than repeat them here: server-log counting of AI fetcher user agents, in the GA4 setup guide, and Search Console’s generative AI performance report. The Search Console one earns a line here for what it does not offer. Google’s help page defines impressions as “how many times links to your site were shown to a user in a generative AI feature on Google Search” and breaks them down by pages, countries, dates and devices. No clicks, click-through rate or query dimension appears anywhere in it, and it covers Google’s own AI Overviews and AI Mode only, so nothing from ChatGPT, Perplexity or Claude reaches it.
A reconciliation you can run in an afternoon
- Pick a window starting after 13 May 2026. Anything crossing that date is not comparable on the native channel.
- Pull both numbers for that window: same property, dates and filters.
- Break your custom channel down by session source, turning “the gap” into a named list of hosts.
- Sort each host into three buckets: counted by both; counted only by you (Perplexity and Claude are the usual candidates); counted only by Google, revealing a host on the unpublished list you had not added.
- Add the third bucket to your regex. Google’s list is a free source-discovery feed.
- Report the union, labelled as a floor, not either figure alone.
Run it monthly. The composition of the gap changes faster than its size, and composition is the interesting part.
Every layer undercounts, and the reason is structural
Every method above counts clicks or fetches, and the dominant behaviour in AI answers produces neither. When an assistant answers completely, including naming and comparing vendors, there is frequently no click: no referrer, no session, no impression anywhere. Google’s own reporting concedes the shape of this, shipping impressions without clicks or CTR.
The residue arrives as Direct sessions and branded searches days later, from people persuaded somewhere you cannot see. We covered building pipeline under those conditions in Zero-Click B2B. The instrument that catches what analytics cannot is dull: a free-text “how did you hear about us?” field on your form.
Where this sits for outbound teams
Measurement tells you which pages the answer engines read and roughly how many people they send. It says nothing about the ten minutes after an enquiry lands, which is where value is won or lost. Zian AI builds autonomous AI sales agents across phone, SMS, email and WhatsApp, with SmartReach AI™ orchestrating channel and timing and PrecisionPitch AI™ split-testing approaches against real outcomes. Zian is in waitlist and partnership beta: if your problem today is analytics configuration rather than follow-up capacity, fix the channel group first.
FAQ
Is the GA4 channel called “AI Assistant” or “AI Assistants”?
Singular. Google’s Default channel group documentation reads “AI Assistant is the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok.” The medium value is ai-assistant and the campaign value is (ai-assistant). Plural spellings are common in third-party write-ups and will not match in a filter or query.
Does the native AI Assistant channel backfill historical data?
Google has not documented it either way. The 13 May 2026 release note and the Default channel group page both say nothing about historical reprocessing. By contrast, Google’s Custom channel groups documentation states that “Custom channel groups can be applied to your reports retroactively.” Until Google says otherwise, avoid year-on-year comparisons on the native channel that cross 13 May 2026.
Does Perplexity traffic appear in GA4’s AI Assistant channel?
Perplexity is not named on either Google page as at 26 August 2026, and Google does not publish the full referrer list, so this cannot be answered from documentation. Test it against your own data: filter sessions to source perplexity.ai and check which default channel group they carry. Keep a custom channel group covering it regardless.
Why do so many AI referrals land in Direct instead?
Because no referrer header arrives. Browsers now default to the strict-origin-when-cross-origin referrer policy, which MDN Web Docs describes this way: “For cross-origin requests send the origin (only) when the protocol security level stays same (HTTPS→HTTPS).” The W3C Referrer Policy specification defines that policy as sending “only the ASCII serialization of the origin of the request client when making cross-origin requests”. Assistant apps and in-app browsers often send nothing, and GA4 classifies a session with no referrer and no UTM parameters as Direct.
Which number should I actually report?
The union of the native channel and your custom channel group for the same post-13 May 2026 window, labelled as a minimum, plus a note that zero-click AI answers are counted by none of these methods. A number presented as a floor survives scrutiny; a number presented as complete does not.
If AI answers are already sending enquiries your way, the bottleneck moves to follow-up speed and persistence. Apply For Partnership to talk to us about autonomous AI agents that work every enquiry across phone, SMS, email and WhatsApp.