AI referral tracking works on two layers. Analytics platforms identify visits arriving from assistants: HubSpot classifies traffic from ChatGPT, Claude, Perplexity, Copilot, Gemini, Meta AI, Mistral, Poe, and Grok as AI Referrals, and OpenAI appends utm_source=chatgpt.com to ChatGPT search referral URLs. Because many assistants and privacy features strip referrers, and because most AI mentions never produce a click at all, referral analytics must be paired with prompt-level citation tracking.
HubSpot has a native AI Referrals traffic source in the traffic analytics tool, so there's no custom channel grouping to build. It recognizes visits from ChatGPT, Claude, Perplexity, Copilot, Gemini, Meta AI, Mistral, Poe, and Grok, and it exposes the specific platform in Drill-Down 1 with campaign parameters in Drill-Down 2, per HubSpot's traffic source documentation.
The part worth wiring up carefully is contact and deal attribution. Original Traffic Source and Latest Traffic Source can both be set to AI Referrals, with the assistant's domain in Drill-Down 1, according to HubSpot's source property documentation. That gives you a list view or report of every contact whose first touch came from an assistant, which is the closest thing to real AI-sourced pipeline reporting available in a CRM right now. We usually build three saved reports at the start of an engagement: AI-referred sessions by platform, contacts created with Original Source of AI Referrals, and deals with any AI Referrals touch in the path.
If you're on GA4 instead, you'll be building a custom channel group with a regex against the known assistant referrer domains, and maintaining that list yourself as new engines appear.
ChatGPT gives you the cleanest signal of the major engines. OpenAI states that referral URLs from ChatGPT search automatically include utm_source=chatgpt.com, which survives most analytics setups and gives you an unambiguous filter.
The rest of the ecosystem is uneven. Some assistants pass a referrer and some pass nothing at all, while plenty of answers get rendered in a way that means the user never leaves the chat window in the first place. Copied links, in-app browsers, and privacy features strip parameters on the way through, so the AI Referrals number you see is a floor, and often a badly understated one.
| Layer | What it sees | What it misses |
|---|---|---|
| HubSpot AI Referrals and UTMs | Clicked visits that arrive with an intact referrer or parameter | Stripped referrers, copied links, zero-click answers |
| Prompt-level citation tracking | Whether you're cited or recommended in the answer itself, per engine | Nothing about what the user did next |
| CRM source properties | Which contacts and deals started with an assistant | Anyone who researched in AI and later typed your URL directly |
Most AI brand exposure never becomes a link at all. Ahrefs found that AI assistants linked to its site in only 28% of brand mentions on average, so for every mention that hands you a measurable click there are two or three that hand you nothing your analytics can hold onto. A buyer can read a recommendation of your product, form an opinion, and show up three weeks later as a direct visit that your dashboard has no memory of.
Tracking the answer itself is what closes that gap. We run a fixed prompt panel across ChatGPT, Claude, Gemini, and Perplexity, rerun it on a schedule, and record whether you appeared, whether you were linked, how you were framed, and who else got named alongside you. That gives you mention rate, citation rate, share of voice, and brand framing accuracy, none of which referral analytics can reach. You can see what the monitoring looks like in practice on our demo of how AEO works, and the same tracking sits inside our AEO Genie tooling.
Neither layer stands on its own. Prompt tracking will tell you the engines know you and recommend you without telling you whether that produced a single dollar, and the CRM data closes that gap once the referred visitors start converting. A program judged on referral analytics alone almost always looks worse than it is, because that's the layer missing most of the exposure.