measurement

Tracking ChatGPT and Perplexity referrals in GA4

GA4 now has a built-in AI Assistant channel. How to check what it catches, build a per-assistant channel group, and read the referrals it can never see.

GA4 now sorts recognized AI assistant referrals into a default channel called AI Assistant, so the first step is to open Traffic acquisition and look for it. The second step is to build a custom channel group that splits that traffic by assistant, because the default channel lumps them together and Google doesn’t publish which referrers it recognizes. The third is to accept what neither can see: visits from assistants that arrive without a referrer land in Direct, and no GA4 setting brings them back.

This post walks through all three in the GA4 interface, with the regex to paste, and ends with the caveats you should put next to the number whenever you report it. It’s one input to the measurement picture in Measuring AI visibility without overstating it. It measures visits, which is the part of AI visibility your analytics can see, not citations.

Does GA4 track ChatGPT traffic automatically?

Yes, for referrers Google recognizes. Google’s default channel group documentation describes AI Assistant as “the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok.” When the referrer matches Google’s list of AI assistants, GA4 sets the session’s medium to ai-assistant and its campaign to (ai-assistant). The channel’s rule is that the medium exactly matches ai-assistant.

Three details on that page change how you read the channel:

  • The list isn’t published. Google names examples, not the full set of recognized referrers. Perplexity isn’t among the examples. It may be recognized; the documentation doesn’t say.
  • Google’s own AI surfaces are excluded. The same page says AI Assistant “excludes Google’s AI Overviews and AI Mode.” Clicks from those land in Organic Search, mixed with ordinary search clicks.
  • The medium changes. Any existing filter, segment, or report that depends on medium values now sees ai-assistant sessions that used to show as referral. Check saved explorations and comparisons that filter on medium.

So the default channel answers “roughly how much assistant traffic do we get,” with an unknown boundary. The next two sections make the boundary visible.

How do you see which assistants send traffic?

Open Reports → Acquisition → Traffic acquisition, change the primary dimension to Session source / medium, and filter the table for ai-assistant. Each row is an assistant’s domain with the medium GA4 assigned. This shows what Google’s list currently catches on your property.

Then check what it misses. Clear the filter and search the same table for the domains you’d expect, one at a time: perplexity, claude, copilot, gemini, chatgpt. Any assistant domain that appears with medium referral instead of ai-assistant isn’t on Google’s list, at least for your data. Write those down; they decide whether you need the custom channel group below.

ChatGPT also tags its own links. OpenAI’s Publishers and Developers FAQ states that ChatGPT adds utm_source=chatgpt.com to referral URLs. That page refused automated fetches when we re-checked it on 11 October 2026, so confirm it in your own data: chatgpt.com should appear as a session source, and in the Landing page + query string dimension you’ll see the parameter on the URLs. The parameter matters because it survives cases where the browser doesn’t send a referrer.

How do you build a custom channel group for AI assistants?

Build it as a copy of the default group with one channel per assistant placed near the top. Google’s custom channel groups documentation gives the path and the constraints:

  • Where: “In Admin, under Data display, click Channel groups.” Copy the default group with Copy to create new rather than starting empty, so everything else keeps its normal channel.
  • How many: standard properties get two custom channel groups (360 properties get five), each with up to 50 channels. Two is not many; agree with whoever else uses the property before spending one.
  • Order matters: “Traffic is included in the first channel whose definition it matches given the current order of channels in the group.” Put the named assistant channels above both Referral and the copied AI Assistant channel, or those rules catch the traffic first. The copied AI Assistant channel can then stay where it is and do the job of the “AI: other” channel below, or you can rename it.
  • It’s retroactive: “Custom channel groups can be applied to your reports retroactively,” so you can read past months through the new group as soon as it’s saved.

Add one channel per assistant you want to see separately, each with a condition on Source using matches regex. A reasonable starting set:

  • AI: ChatGPT — Source matches regex .*(chatgpt\.com|chat\.openai\.com).*
  • AI: Perplexity — Source matches regex .*perplexity\.ai.*
  • AI: Gemini — Source matches regex .*gemini\.google\.com.*
  • AI: Copilot — Source matches regex .*copilot\.microsoft\.com.*
  • AI: Claude — Source matches regex .*claude\.ai.*
  • AI: other — Medium exactly matches ai-assistant (or keep the copied AI Assistant channel, which uses the same rule)

The leading and trailing .* make each pattern work whether the field is compared as a full match or a partial one, and also catch subdomains. The backslashes make each dot literal.

These domains are a starting point, not a documented list. Take the actual source values from the Traffic acquisition check above and adjust the patterns to match what your property records. The last channel catches anything Google recognizes as an assistant that your named channels missed, so new assistants show up as “AI: other” instead of disappearing into a total.

Once saved, select the new group from the primary dimension menu in Traffic acquisition and you have per-assistant sessions, engagement, and key events.

How do you see which pages assistants send people to?

Use a free-form exploration. Go to Explore → Free form, then:

  1. Dimensions: add Session source, Landing page + query string, and your new custom channel group.
  2. Metrics: add Sessions, Engaged sessions, and Key events.
  3. Rows: Landing page + query string. Columns: your custom channel group.
  4. Filter: custom channel group begins with AI:.

The result is a table of the pages each assistant sends visitors to. That list is useful for more than reporting. Those landing pages are, by direct evidence, pages that assistants link to. Check them first when you audit page-side readiness: they’re the pages where a blocked crawler or a broken template costs you visits you already have. Prioritizing AEO fixes uses exactly this kind of signal in its page-value column.

To export, use the share/export menu at the top of the exploration and download to a spreadsheet. No API or terminal is needed for any of this.

What does GA4 miss?

GA4 misses any assistant visit that arrives without a referrer or a UTM tag, and it can’t separate AI Overviews clicks from organic search. Those are structural limits, not settings you’ve got wrong.

Missing referrers. GA4 attributes a session to a source using the referrer the browser sends, or tags on the URL. When neither is present, the session is Direct. That happens when a user copies a link out of an answer and pastes it, when a link opens from a desktop or mobile app that doesn’t pass a referrer, and when the referring page’s policy strips it. None of the assistant vendors documents how often their apps send referrers, so there’s no published way to estimate the gap. Treat AI Assistant sessions as a floor.

Google’s AI features. Clicks from AI Overviews and AI Mode count as Organic Search. Google’s AI features documentation says sites appearing in those features are reported in Search Console’s Performance report “within the ‘Web’ search type,” not as a separate line. GA4 can’t split them; The Search Console generative AI report, read honestly covers what Search Console can and can’t tell you about that slice.

Answers without clicks. An answer that names or quotes your page without the user clicking leaves no trace in GA4. Referral sessions measure visits, which is a subset of being cited. What prompt tracking can and can’t tell you covers the method that tries to observe the answers themselves, and its limits.

Bot traffic. Assistant crawlers fetching your pages don’t run your analytics tag, so crawler activity doesn’t appear in GA4. That’s in your server or CDN logs; Reading server logs for AI bot traffic covers that source.

Visit typeWhere it shows in GA4
Click from a recognized assistant, referrer sentAI Assistant channel
Click from an assistant Google doesn’t recognizeReferral (until your custom group catches it)
ChatGPT link with utm_source=chatgpt.com, no referrerAttributed by the UTM tag
Copy-paste or in-app click, no referrer, no tagDirect
Click from AI Overviews or AI ModeOrganic Search
Crawler fetchNot in GA4

How should you report the number?

Report assistant sessions as a count with the date range and the channel definition next to it, and say in the same sentence that it’s a floor. “1,240 sessions from AI assistants in October, by GA4’s AI Assistant channel plus our custom rules; visits without a referrer appear as Direct and aren’t included” (an illustrative figure) is a sentence a client or executive can repeat without overstating it.

Avoid percentage-of-traffic framing unless you give the denominator and it’s stable. Small absolute numbers make large percentage swings, and a doubling from a low base reads as a trend when it may be noise.

Pair the number with page-side work you can verify. Referral sessions tell you which pages assistants send people to; a crawl tells you whether those pages are reachable and readable, and a re-crawl after fixes shows whether they changed. Crawl, fix, re-crawl covers that loop, and How to audit a site for AEO and GEO covers the audit around it.

GEOBee doesn’t read analytics. It covers the other half: whether the landing pages your exploration surfaces can be reached by the AI crawlers and lifted cleanly once fetched. To audit those pages on your own machine, download GEOBee and start the 7-day trial.

On this page