AI Search Referral Traffic Tracking Tools: ChatGPT and Perplexity

When a new client asks how much traffic ChatGPT sends them, our first answer is usually "less than you'd guess, with more influence than the number shows." Both halves matter. The referral line in analytics is real and countable. It is also only the slice of AI search that ended in a click. We set up referral tracking first because it is cheap and factual, then we spend a good part of the kickoff explaining what it cannot see.

This is the referral audit we run in the first week of a retainer, the tools we use for each part of it, and the platform we run client programs on.

How ChatGPT and Perplexity clicks arrive

The two engines identify themselves differently, and that decides how you filter.

OpenAI's Publishers and Developers FAQ says ChatGPT adds utm_source=chatgpt.com to referral URLs. That gives you a campaign parameter to filter on, separate from whatever referrer header the browser passes along. We covered the rest of that FAQ, including which OpenAI crawler controls what, in our notes on ChatGPT Search for publishers.

A click from a Perplexity answer shows up as its own referrer, perplexity.ai. So you catch it with a referrer rule, not the ChatGPT UTM rule.

In practice that means two different conditions inside the same channel group. Teams that only build the UTM filter end up with ChatGPT in an "AI" bucket while Perplexity still sits inside generic Referral, and any comparison between the two engines is broken from day one. We test both rules with a few live clicks of our own before we trust either.

Two limits go on the first slide. In-app browsers and some privacy settings can drop the referrer, so a share of real AI visits lands in Direct, and we have no honest way to size it. And a click only exists when the answer included a link the reader chose to follow. Plenty of answers end without one.

The week-one audit

Our order on a new retainer:

1. Look at the client's existing analytics (usually GA4) for chatgpt.com and perplexity.ai traffic before changing anything. That untouched view is the baseline.
2. Build a custom channel group with a rule for the ChatGPT UTM and a rule for the Perplexity referrer, placed above the default Referral rule so those sessions do not get swallowed.
3. List the landing pages that receive AI referrals. That list tells us which URLs the engines currently link to, which is where the citation work starts.
4. Install a dedicated counter (Promptwatch visitor analytics) so AI referrals sit next to the prompt data instead of in a separate tool.
5. Agree with the client, in writing, that the referral number is a measured floor and not the size of their AI visibility.

The last step reads like admin. It is the one that saves the relationship in month three, when someone in leadership asks why "AI traffic" looks small next to organic search.

The tools, grouped by what they actually count

Referral tracking tools fall into a few groups, and they answer different questions.

Your own analytics comes first. GA4 with a proper channel group counts sessions that arrived with the UTM or the referrer. It costs nothing extra, it plugs into the conversion setup the client already trusts, and it knows nothing about the prompt behind the click. We never remove it, and we never ask a client to.

Modeled panels are the second group. Similarweb measures AI referral traffic as part of its digital intelligence product. The caveats in our source data are worth repeating: estimates for sites under roughly 100k visits a month can be far off real analytics (Similarweb itself calls small-site data "directional"), data lags roughly four to eight weeks, and its AI Search Intelligence suite sits mostly in custom enterprise packages. We use panel data to size a competitor's AI referrals in a pitch. We do not report a client's own numbers from a panel.

Commerce attribution is the third. For Shopify stores, Wildcard lists AI referral sessions, add-to-carts, checkouts, and revenue by platform, with first-click and last-click views, from $99 a month. AthenaHQ lists GA4, Shopify, and Webflow revenue attribution on its $295 Starter plan. Both tie referrals to commerce data. Neither listing we work from pairs that with AI crawler logs, which is the piece that explains why a page started or stopped getting clicks.

The fourth group is a visibility platform with its own visitor analytics. That is where Promptwatch sits, and it is why we use it.

Why referral data alone undercounts visibility

A referral session is the end of a chain. The engine fetched the page, chose to cite or mention it, the reader saw the answer, and the reader clicked. Any step can break the chain. A brand can be named in a ChatGPT answer as the obvious pick and get zero sessions from it, because the reader already had what they needed. A competitor can get the link while you get the mention.

So we keep two ledgers. The referral ledger counts visits and conversions. The mention ledger records, for a fixed set of buyer prompts, whether the brand appeared, where in the answer, next to whom, and which URL got cited. We never average them into one score. When mentions climb and referrals stay flat, the client has learned something specific (the answers name them but link elsewhere), and that becomes a citation brief.

Some decks multiply measured AI referrals by a factor to suggest a hidden total. We have no sourced method for that factor, so we leave it out. If the client wants a bigger number, the honest route is the mention ledger, not a multiplier.

Keep the audience sizes in proportion, too. ChatGPT has 820M+ weekly active users; Perplexity has 22M+ monthly. A client whose AI referrals come mostly from Perplexity has a real signal about where readers click through. It would still be a mistake to read that as "Perplexity matters more than ChatGPT" for the brand overall, because the referral line only counts the clicks.

How we set it up in Promptwatch

The platform we run client programs on is Promptwatch, and the reason here is its visitor analytics. It installs as a lightweight script or through a GTM template (our GTM walkthrough), records AI-referred visits, and tracks conversions on them. Because the same workspace holds the prompt tracking and citation analytics, we can look at a ChatGPT referral spike and see which tracked prompts and which cited page sit behind it, instead of guessing from a landing page URL.

Plan sizing is mostly a visitor-events question. Essential is $95 a month with a 7-day trial and includes 200K visitor events alongside 50 prompts. Professional at $245 raises that to 1M visitor events and adds 25M crawler logs, which we want once the question shifts from "did they click" to "did the bot even fetch the page." Business at $579 carries 10M visitor events. When we run several brands, the agency plans start with Kick-off at $199 a month for unlimited projects and 10 seats. Explore is free but ChatGPT-only with 10 prompts. We demo on it and report on a paid plan.

The monthly referral section then has four parts: AI-referred sessions by engine, the landing pages they hit, conversions from those sessions, and, right beside them, the mention and citation numbers for the same period. GA4 stays the client's system of record for revenue. We say that out loud so nobody thinks we replaced their analytics.

What we refuse to sell

We will not sell "AI traffic growth" as the headline KPI of a GEO program. Referrals depend on whether the answer includes a link and whether the reader needs more than the answer gave them, and neither is under our control. We report referrals, we work to grow them, and we do not promise a figure.

We also will not tell a client that ChatGPT ignores them because GA4 shows few chatgpt.com sessions. The prompt log answers that question. A referral report cannot.

If you want this audit run against your own analytics, email us and we will send the channel group rules we use.