Guides

Reading GA4 for AI referrals: what you can and can't attribute

Some AI assistants pass a referrer and many don't. How to measure the visits you can identify without inventing a total.

"How much traffic do we get from ChatGPT?" is a fair question and an awkward one, because the honest answer is "at least this much."

#Why it's incomplete

When someone follows a link in an assistant, whether your analytics sees a referrer depends on the assistant, the app or browser, and whether the click opened in a new context. Some visits arrive with a recognisable referrer. Some arrive tagged with a campaign parameter the provider appends. Others arrive as direct traffic with nothing to identify them.

So any "AI traffic" figure falls into three buckets:

  1. Identified: a referrer or parameter points to a known assistant.
  2. Probably AI, unidentifiable: a spike of direct traffic to a deep page that no one links to. Interesting, but not provable.
  3. Unknown: everything else.

Only the first bucket should appear in a report as a number. The others are context.

#Setting up the identified bucket

In GA4, build a custom channel group or a segment that matches the source/referrer values you see for assistants. Because those values change, treat the rule set as maintained, not set-and-forget:

  • Review the list of sources quarterly.
  • Keep the rule list documented so a colleague can reproduce the numbers.
  • Report the rule version with any figure you publish.

#Measure outcomes, not just sessions

Sessions say little on their own. For each identified assistant, look at:

  • Landing pages. Which pages do AI visitors land on? Often documentation and comparison pages, not the homepage.
  • Engagement. Pages per session, time on page, scroll — compared with other channels.
  • Key events. Sign-ups, demo requests, purchases.

Small volumes need care: a handful of sessions can't support a conversion-rate claim.

#Crawlers aren't visits

Server logs show bots too: search crawlers, training crawlers and user-triggered fetches. They aren't audience, and they don't appear in GA4 as sessions. If you analyse logs, keep these categories separate from human referrals.

#Pair it with visibility data

The useful question isn't "how many visits?" but "which answers lead to visits?" Compare your identified referral landing pages with the prompts where you're cited. You'll often find a few pages doing most of the work — and that's where to invest first.

Citeroot doesn't connect to GA4 yet (a Referral Radar module is planned — join the waitlist). Until then, a quarterly review of identified sources, landing pages and outcomes, set beside your Citeroot answer data, goes a long way.

analyticsga4referrals

See what AI says about you today.

Add your site and a few questions your buyers ask. Get your first baseline in an afternoon — 14-day free trial · no card.