Tracking ChatGPT and Perplexity Referrals Without GA4

Visitors arriving from ChatGPT, Perplexity, Copilot, and Claude are a real referral source for clinic websites now, and you should be able to see them as one. The standard advice for doing that is to build a custom channel group in GA4, which is the wrong fix for a healthcare site, because a clinic should not be running GA4 on patient-facing pages in the first place. Curve records the referring hostname on its own servers, under a BAA, so chatgpt.com and perplexity.ai show up as sources in your reports, and Curve Analyst lets you ask how that traffic is doing in plain language and get a chart back.

TL;DR

• AI assistants pass a referrer hostname (chatgpt.com, perplexity.ai, copilot.microsoft.com, and others) on most clicks. Some clicks arrive with no referrer and get counted as direct, so whatever you measure is a floor.
• Most analytics tools lump these hostnames under a generic "referral" channel, which is why teams believe they have no AI traffic when they do.
• The usual fix, a custom channel group in GA4, requires GA4, and Google does not sign a BAA for GA4. That rules it out for pages where patients browse services or book.
• Volume is the least interesting number. Landing pages, conversion rate against organic search, and which pages get cited tell you what to do next.
• Getting cited by an assistant is an optimization problem. Measuring what happens after the click is an analytics problem. Solve them separately.
• Curve records the referring hostname on its own servers, so each assistant appears as a source, and Curve Analyst answers questions about them from your own data, with no GA4 involved.

How do AI assistant referrals show up in your analytics?

They show up as a referrer hostname, the same way a link from a local news site does. When someone asks ChatGPT which clinics near them do a certain procedure and clicks the link in the answer, the browser sends a request to your site with the referrer set to chatgpt.com. Perplexity sends perplexity.ai. Microsoft Copilot sends copilot.microsoft.com. Claude sends claude.ai when a link is clicked from a conversation. Gemini sends gemini.google.com.

That is the tidy case. The messy case is the click that arrives with no referrer at all. It happens when the assistant runs inside a native app rather than a browser, when the link opens through an intermediate redirect that drops the header, or when a privacy setting on the visitor's device suppresses it. Those visits land in your reports as direct traffic, indistinguishable from someone typing your URL. Treat every AI referral count you see as a floor.

What a session from an AI assistant looks like

The other tell is the landing page. Search sends most people to your homepage or a service page you have optimized for a query. An assistant sends people to whatever page it cited in its answer, which is frequently a blog post, an FAQ, a pricing explainer, or a page about insurance and financing. The assistant already did the comparison shopping, so the visitor arrives further along and with a narrower question: do you take my plan, and how soon can I be seen.

Why do most tools bury AI referrals under "referral"?

Because the channel grouping rules were written before assistants sent traffic. Every analytics product ships a default set of channel rules: paid search, organic search, social, email, direct, referral. The rules assign a session to a channel by matching the referrer hostname, the medium, or a click ID against a list. Any hostname the list does not recognize falls into the catch-all: referral. chatgpt.com is on almost nobody's default list, so it lands in the same bucket as a link from a Yelp page or a supplier's website.

You can still find it. Open the referral channel, break it down by source, and scroll. Nobody does that on a Monday morning while trying to explain last month's numbers. Channel reports get read at the channel level, so if the channel says "referral" and the number is small, the conversation moves on and the assistants recommending your clinic never surface.

The common fix is to write your own channel definition. In GA4 that means a custom channel group with a rule that matches a list of AI hostnames and labels them "AI search" or similar. It works, and we wrote about the mechanics ourselves in our guide to measuring AI search traffic in healthcare. The catch is what the fix sits on top of.

The GA4 problem for a clinic website

Google does not sign a BAA for Google Analytics 4. Under HIPAA, a vendor that handles protected health information on your behalf is a business associate and needs a BAA. On a clinic site, the URL a visitor lands on, the pages they view, the form they submit, and the time they spend on each can all say something about that person's health, and GA4 collects those alongside identifiers like the client ID and IP address. The FTC has also pursued health data disclosures to ad platforms on unfairness grounds, as in the Hims & Hers complaint, so the exposure is not limited to HIPAA's reach.

A custom channel group does nothing about any of that. It relabels data that was already collected in a way you cannot stand behind. Building better AI reporting on GA4 for a clinic is tidying a room you should be moving out of. If you want the longer argument, read whether GA4 is HIPAA compliant and what the alternatives are.

What should you look at beyond raw AI traffic volume?

Volume is the number everybody asks for first and the one that tells you the least. A count of sessions from chatgpt.com says an assistant is recommending you. It does not say what for, or whether those people became patients. Three other views do.

Landing pages

Which pages do AI visitors arrive on? This is the closest thing you have to a list of what the assistants are citing. If the answer is your blog post on recovery timelines rather than your service page, you have learned something about what the assistant considers worth pointing to, and something about where your booking path needs to be visible. A landing page that receives AI referrals and has no clear next step is a leak you can fix this week.

Conversion rate against search

Put AI referrals next to organic search and paid search, and compare the rate at which each becomes a booked appointment or a submitted form. Do not compare raw counts, because search will win on count for a long time. Compare the rate, and compare it per goal. Expect the pattern to differ by specialty and by landing page, which is why your own numbers matter more than any benchmark. If the rate is higher than search, you have a reason to invest in being cited more. If it is lower, look at the landing pages before you conclude the traffic is low quality; it may be arriving on pages with no way to book.

Which sources, over time

The assistants are not interchangeable. One may send you traffic that converts and another may send traffic that leaves immediately. Track them as separate sources under a shared channel, and look at the split month over month. That is enough to spot when one assistant starts or stops citing you.

Is getting cited by ChatGPT the same problem as measuring it?

No, and conflating them is how teams end up with a dashboard and no plan. Being cited is an optimization problem. It depends on whether your pages answer the question an assistant is trying to answer, whether the content is structured so the assistant can extract it, whether your practice appears in the directories and reviews the assistant draws on, and whether the page is accessible to the crawlers those assistants use. That discipline has a name now, answer engine optimization, and it has its own playbook. There is a guide to it in how healthcare practices get cited by ChatGPT, Perplexity, and Claude, and the broader picture in what AI search means for healthcare marketing.

Measuring is an analytics problem. It starts after the click and asks what the visitor did. You can be cited constantly and measure nothing, or measure perfectly and be cited never. The two feed each other: measurement tells you which pages are earning citations and converting, and that tells you what to write more of. But they are different work, usually done by different people, and the measurement half is the one you can fix this month.

How does Curve handle AI referrals without GA4?

Curve records the referrer hostname on its own servers when the pageview arrives, so each assistant shows up as its own row in the sources report: chatgpt.com, perplexity.ai, copilot.microsoft.com, and so on, next to google and facebook. At the channel level they still count as referral today, so for now the sources report and a source filter are how you see them as a group, and that is enough to answer every question below.

The collection path is what makes this usable on a clinic site. Events go to Curve's US-hosted infrastructure first, server-side, where anything that should not leave is stripped before conversions are forwarded to ad platforms. Curve signs a BAA with every customer on every plan. No GA4 tag is involved. The same referrer source is carried into goals and funnels, which is what lets you compare AI visitors against search visitors on conversion rate instead of on session count.

Where Analyst fits

Analyst is the chat inside the Curve dashboard that answers questions about your own account's data. Ask how AI search traffic looks this month and you get a trend line and a source breakdown, with the prior period alongside. Ask which pages AI visitors land on and whether they convert, and you get a table. It draws on the same three areas as the dashboard: website analytics, goals and funnels, and campaign reporting. Every number is the result of a query against your account; it never estimates, and if a metric is empty or still syncing it says so before giving a figure.

It runs on Claude through Amazon Bedrock, inside infrastructure covered by Curve's BAA with AWS. Your data is not exported to a separate AI product and is not used to train models. We built it read-only, so it cannot change tracking or destinations, and it does not look up individuals or answer questions about a specific person. It keeps the date range and filters of the screen you opened it from, so open it from the sources report and its answers stay inside that window. There is a longer explanation in what Curve Analyst is and how it works.

What to ask Curve Analyst

Start with these three, then follow the thread. The first gives you the headline for the Monday meeting. The second tells you what to do about it. The third puts AI traffic in context against everything else that brings people to the site.

• "How is AI search traffic looking this month compared to last month?"
• "Which pages do visitors from AI assistants land on, and do they convert?"
• "Show me my top sources over the last 30 days with conversion rate for each."

Analyst also offers suggested questions when you open it. The launch announcement walks through what it can and cannot answer: Introducing Curve Analyst.

Frequently asked questions

Does AI assistant traffic show up as direct?

Some of it does. When the click comes from a native app, through a redirect that drops the referrer, or from a device with a privacy setting that suppresses it, there is no hostname to record and the session is counted as direct. Any AI referral figure is therefore a floor. Direct traffic that lands on deep content pages rather than your homepage is often assistant traffic that lost its referrer on the way in.

Can I track ChatGPT referrals without Google Analytics at all?

Yes. Any analytics tool that records the referrer hostname can separate them; the question is whether it does so by default and whether you can run it on a clinic website. Curve records the referrer on its own servers, keeps the hostname as the source, and signs a BAA, so the report is there when you log in.

Why does my AI referral traffic look small?

Because most of it is hidden in the referral bucket or in direct, and because assistants send fewer, more specific visitors than search does. Small is not the same as unimportant. Compare conversion rate against search before you decide how much attention it deserves.

Can Curve Analyst tell me which pages ChatGPT is citing?

It can tell you which pages visitors from ChatGPT land on, which is the closest proxy available from your own data. It cannot see inside ChatGPT's answers, so a page that gets cited but never clicked will not appear. For the citation side you need answer engine optimization work and the tools built for it.

Does Curve Analyst answer questions about individual visitors?

No. It does not look up individuals and does not answer questions at the person level. It works in aggregates: sessions, pages, sources, goals, and campaigns. That limit applies to AI referral questions the same as any other.

If you want to see what AI assistants are sending to your site without putting GA4 back on it, book a demo with Curve and ask Analyst the three questions above against your own account.