Tracking AI assistant referral traffic in GA4

Set this up before the work starts

Without a baseline, six months of AI search work produces numbers with nothing to compare against. The setup takes under an hour. Understanding what the numbers mean is the harder half, and it is where this guide spends most of its time.

What Google Analytics now does by default

DOCUMENTED Google Analytics 4 defines a default channel called AI Assistant: “the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok”. Its rule is that the medium exactly matches ai-assistant; when a session’s referrer matches a recognised AI assistant source, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant). So in a current property, assistant referrals that carry a referrer are grouped for you under Reports, then Acquisition, then Traffic acquisition, with the primary dimension set to Session default channel group.

Two caveats. The list of sources Google recognises is Google’s and changes; a new assistant may land in Referral until it is added. And the channel only sees sessions that arrive with a referrer at all, which is the floor problem below.

The hosts to watch

Whether or not the default channel catches them, these are the hostnames assistant referrals carry when a user clicks an attribution link:

Review the list quarterly. New hosts appear as products change.

Finding them in GA4

Reports, then Acquisition, then Traffic acquisition. Set the primary dimension to Session source and filter for one of the hosts above, or set it to Session default channel group and look for AI Assistant. If nothing appears, check two things before concluding there is none: the date range (these numbers are small and a seven-day window may genuinely be zero) and the consent configuration (a banner that blocks analytics until acceptance hides every non-consenting session, and small channels lose the largest share).

Building a reusable view

• Exploration segment. Explore, blank exploration, new segment scoped to Session, condition Session source matches regex chatgpt\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai. Name it “AI assistants” and apply it to a free-form table with landing page as the row and sessions, engaged sessions and key events as values.
• Custom channel group. Admin, Data display, Channel groups. Create one based on the default, add a channel “AI assistants” positioned above Referral, using the same regex against Source. This catches hosts the default AI Assistant rule does not yet recognise and makes the channel available in standard reports.

The landing-page dimension is the useful one. It shows which pages assistants send people to, which often contradicts expectations.

Why the numbers look small

INFERRED Expect single or double digits a month for most Perth businesses, even where the work has an effect, for four reasons none of which better configuration fixes:

1. Most AI-influenced journeys have no click. The assistant names three providers; the user reads, forms an impression and searches the brand later. That session arrives as branded organic or direct.
2. Attribution links are one option among several, and people read the answer far more often than they follow its sources.
3. Some sessions carry no referrer. App and desktop clients can strip it, and those sessions arrive as direct.
4. The influence is often several steps back: a conversation in week one shapes a shortlist, the enquiry arrives in week five through a branded search.

Treat the referral count as a floor and its movement as directional.

Things that quietly break the tracking

• Consent mode. Sessions from users who never accept are missing entirely; small channels lose most.
• Internal traffic filters. A stale IP range from a previous office removes real sessions.
• Cross-domain gaps. An enquiry form on a subdomain not configured for cross-domain measurement restarts the session at the handover, and the original referrer is lost right before the conversion.
• Referral exclusions. An assistant host added to the referral exclusion list turns those sessions into direct. Check nobody did this to tidy the report.

The most common cause of a zero is one of these, not an absence of visitors.

What assisted conversions can and cannot prove

Attribution paths show whether an AI-sourced session participated anywhere in a converting journey, which is the strongest analytics evidence available for this channel. They cannot prove the interaction caused the conversion, and they cannot see impressions: the largest part of AI influence is a name read without a click, invisible to analytics by construction. The honest framing is that referral and assisted figures are a visible fraction of a larger effect with the fraction unknown. Anyone quoting a precise multiplier is inventing it.

Filling the gap with proxy signals

• Branded search in Search Console. Impressions and clicks for queries containing the business name, read as direction over months. DOCUMENTED Google’s AI feature clicks also sit inside Search Console’s Web totals, so branded movement is the closest available proxy for shortlist inclusion on Google.
• Direct sessions to deep pages. A visitor landing directly on a specific guide or service page was usually given the URL by something. Some of that is email; a steady rise alongside AI work is worth noting.
• An enquiry-form question. “How did you hear about us?” with an explicit AI assistant option. Self-reported and imperfect, and still better data than the platform can natively provide.

What to report, and how often

Monthly: sessions and key events from the AI assistants channel with the trend, top landing pages from that channel, branded organic clicks as a supporting line, and work completed. Quarterly: prompt-set movement from the repeated-run baseline. State the under-count in the report itself, because the person reading it in six months will not remember the conversation.

What is documented and what is inferred

• Documented: GA4’s AI Assistant default channel and its rule; Search Console’s treatment of AI feature traffic.
• Inferred: the reasons referral counts are a floor; the configuration failures; the proxy signals.
• Not established: any Perth-specific referral volume.