GA4 now has a native AI Assistant channel, but it identifies AI visits by referrer alone, and the list of assistants it recognises may not include Perplexity. To see ChatGPT, Gemini and Perplexity reliably, build a custom channel group with an anchored regex placed above Referral, add an OR condition that also captures the native channel, and validate the two against each other. Then read whatever number you get as a floor, because a meaningful share of AI-driven visits never carries a referrer at all.
That last point matters more after October 2026. Both ChatGPT and Gemini reportedly changed how they tag outbound links within a month of each other, and neither change was announced. If you report AI traffic to a board or a client, the method below gives you a stable definition and a way to tell a tracking change from a change in demand.
What GA4 now tracks natively, and where it stops
The AI Assistant channel launched on 13 May 2026
Google added an AI Assistant channel to GA4's default channel group on 13 May 2026. Qualifying sessions get the medium ai-assistant, the channel group AI Assistant and the campaign (ai-assistant), all assigned automatically. According to a write-up on nicelookingdata.com, Google's Default Channel Group documentation names five recognised sources: ChatGPT, Gemini, Deepseek, Copilot and Grok.
The rule behind it is blunt. GA4 assigns the channel in two cases only: the medium is ai-assistant, or the referrer matches its list of known assistants. It ignores landing pages, behaviour and everything else. No referrer and no tag means no AI attribution.
One more claim is doing the rounds: that the native channel is forward-only, so sessions before May 2026 won't be reclassified. We haven't seen that independently confirmed, but it fits how default channel definitions usually behave. It's one more reason to keep your own definition.
Why Perplexity and other assistants need checking in your own property
Published sources disagree on the list. One says Claude is covered. Another gives the five above, with neither Claude nor Perplexity. At least one guide says Perplexity isn't in Google's channel description, so its visits may still sit under Referral.
Don't trust any blog's list, this one included. Open Reports, Acquisition, Traffic acquisition, set the primary dimension to Session source / medium and search for perplexity. If you see perplexity.ai / referral, the native channel isn't claiming it in your property. Check Google's Default Channel Group documentation too, because Google can edit that list without notice.
Build a custom AI channel group anyway
Even if the native channel covered every assistant today, we'd still build a custom group. You get one definition of "AI traffic" that you control. It covers Perplexity and anything Google hasn't named. And because custom channel groups are evaluated at query time, it applies to historical date ranges. The trade-off is maintenance: you own a hostname list and have to review it whenever a new assistant gains traction.
The path is Admin, Data display, Channel groups. Copy the default channel group rather than starting from scratch, so you keep every standard channel and only add one.
Write an anchored regex and place it above Referral
Add a new channel, give it an unambiguous name like "AI assistants", and set the condition to Session source matches regex. This is the pattern we start from:
^(gemini|(.*\.)?(chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|deepseek\.com|grok\.com|meta\.ai|you\.com))$Several choices in there are deliberate:
- It's anchored. The
^and$, combined with the optional subdomain group, matchchatgpt.comandwww.perplexity.aibut not a lookalike such asnotchatgpt.com. - Dots are escaped. An unescaped
.matches any character, which is a quiet way to over-match. chat.openai.comis included because older ChatGPT links still use it.- Bare
geminiis included to catch sessions from the period when Gemini reportedly tagged links withutm_source=gemini(more on that below). - There's no
bingand no barecopilot. We've seen published segments that includebing, which sweeps ordinary Bing search referrals into "AI". Unanchored terms like that inflate your numbers in a way that looks plausible, which is the worst kind of error.
Order is the step people miss. GA4 evaluates channels top to bottom and the first match wins. Put your AI channel below Referral and Referral claims the sessions first, leaving your new channel empty. Drag it above.
Test before you trust
Before saving, paste the source values you actually see in Traffic acquisition into a regex tester and confirm each one matches or doesn't. Hostname lists drift, and a typo here fails silently.
Add the OR condition for the native channel
Add a second condition group joined with OR: Default channel group exactly matches AI Assistant. Your channel becomes the union of Google's definition and yours, whether you're building fresh or replacing a homemade channel from before May 2026. When Google adds a source you haven't listed, or a visit arrives with medium ai-assistant, you still catch it.
Validate against the native channel
Go to Traffic acquisition, switch the primary dimension to Session custom channel group, and compare your AI channel with the native AI Assistant channel over the same date range. Add Session source / medium as a secondary dimension to see each platform row by row.
For any period after 13 May, your channel should be equal to or larger than the native one. The difference should be explainable source by source, typically Perplexity or Claude. If yours is smaller, something is ordered wrongly or the OR condition didn't save. If it's dramatically larger, look for an over-matching pattern.
What broke in October 2026
Both episodes below have already happened, and neither was documented by the platform at the time of writing. Treat them as reported, observed behaviour.
ChatGPT and the reported GPT-6 tagging gap
Search Engine Roundtable reported on 14 October 2026 that utm_source=chatgpt.com had disappeared from ChatGPT links with the release of GPT-6 about a week earlier. An Adobe employee's LinkedIn post described the mechanism: GPT-6 opened links with a window.open call using noopener,noreferrer and dropped the UTM. With no referrer and no tag, GA4 records the visit as (direct) / (none).
The scale varied by site. Consultant Natzir Turrado's analysis covering 7 to 16 October reported 68% to 74% of ChatGPT-channel visits lost in server logs, against 20%, 46% and 60% in GA4 depending on the site. These are small-sample figures, not a population statistic, but the gap was big enough to show up in a monthly report.
Whether it's fixed is contested. One source says OpenAI confirmed a bug and that almost all links have carried the tag again since 16 October. Another describes the change as unconfirmed by OpenAI, and a third notes that the tag returned but the referrer may not have. OpenAI's own help page says ChatGPT automatically includes utm_source=chatgpt.com in referral URLs. That's the official position, though the page is undated.
Gemini's reported tags, added and then dropped
Around 30 September to 1 October, Gemini quietly began adding UTM parameters to outgoing links. A Reddit user spotted them, John Mueller said in the thread that he could see them too, and Google didn't document the change. The reported value was utm_source=gemini.
That mattered, because the Gemini referrer mostly survives only on desktop web. Search Engine Journal reported referrer pass-through from the Gemini mobile app at around 9 percent, with Android assistant invocations stripping it entirely. A UTM survives the in-app webview where a referrer doesn't.
It didn't last. Turrado asked on 27 October whether others were seeing the tag gone, and Search Engine Roundtable reported on 30 October that Gemini had stopped attaching UTM parameters to cited URLs. Google hasn't explained either change.
Keep Gemini the assistant separate from Google Search's AI features. A gemini.google.com source is the assistant, not AI Mode, and there was no confirmed tagging for AI Overviews or AI Mode clicks as of early October. If Search is the surface you care about, that's a different measurement problem, covered in our guide to tracking your visibility in Google AI Overviews.
Treat your AI numbers as a floor, not a total
Everything above rests on referrers or tags, and both are fragile. Claude doesn't tag its links at all. ChatGPT's mobile apps rarely pass a referrer. Turrado reports Gemini Android traffic landing in Direct almost without exception, and the Gemini iOS app sending no referrer. Platforms can change any of this without notice, as October showed twice.
Report AI traffic as "at least this much", trend it against itself, and never put it side by side with organic search as if the two were measured the same way.
Organic search attribution is mature and fairly stable. AI attribution is neither. Put them in one chart and the reader concludes AI is small, when part of what they're seeing is a measurement gap.
Where the missing sessions hide in Direct
The untracked share lands in Direct. To find it, compare Direct sessions by landing page across the dates that matter: before and after 30 September, 7 October and 30 October. Watch the deep pages assistants tend to cite, such as guides, comparison pages and documentation. A rise in Direct to a long-tail article nobody bookmarks is worth investigating.
Be careful with conclusions. A Direct shift that coincides with a reported tagging change is a lead, not proof. Confirm it in server logs before you tell anyone the AI channel "really" grew. A "How did you hear about us?" field on forms is a cheap first-party fallback that no platform change can remove.
Get plain-English guides like this in your inbox.
One short email a month. WordPress, Shopify, SEO, no fluff. Unsubscribe in one click.
We never share your email.
Annotate, baseline and triangulate with server logs
October made one thing obvious: in a dashboard, a tracking change looks exactly like a demand change. The fix is process, not a cleverer regex.
Annotations and a dated baseline
Add GA4 annotations for 13 May (native channel), around 30 September (Gemini tags reported), 7 to 16 October (ChatGPT gap) and 30 October (Gemini tags reportedly dropped). Then record a dated baseline: weekly sessions for your AI channel, the Gemini rows specifically, and Direct to your most-cited pages. Custom channel groups apply at query time, so you can build that baseline backwards from before May.
Tell stakeholders before the quarterly deck, not after. A short note saying AI figures dipped for a week in October because of a reported platform change saves a much longer conversation about falling demand that never happened. We build this kind of reporting discipline into our digital marketing and analytics support.
Landing page query strings and server logs
In GA4, use the Landing page + query string dimension and filter for utm_source=chatgpt.com or utm_source=gemini. That shows tagged arrivals even when the referrer was stripped, and it's a quick way to check whether a platform is tagging today.
Server logs are the stronger evidence. Every October finding above came from logs, not from GA4. Logs record the request and query string whether or not your analytics script fired or consent was given. If your GA4 data is already patchy from consent and blockers, read our piece on server-side GA4 and conversion tracking, which covers how to close that wider gap.
How we would set this up for your team
For most teams, we'd work in this order:
- Check Google's live Default Channel Group documentation and your own Traffic acquisition report for Perplexity and Claude.
- Copy the default channel group, add an AI channel with an anchored regex, place it above Referral and add the OR condition for AI Assistant.
- Validate against the native channel, source by source.
- Add annotations for the known dates and record a dated baseline, including Direct to cited pages.
- Set up a monthly check of landing page query strings and server logs for
chatgpt.comandgeminitags. - Report AI traffic as a floor, trended against itself, with a one-line caveat stakeholders actually read.
Measurement is half the job. Once you can see the traffic, the next question is how to earn more of it, which is where generative engine optimisation comes in. If you'd like a second pair of eyes on your property, get in touch about your GA4 setup and we'll reply within four hours during UK business hours.
