Stop bragging about ChatGPT mentions you can’t tie to trials

Two separate tracking guides landed within three weeks of each other, and both said the same thing: AI assistants are sending real traffic, and GA4 is hiding a chunk of it. Ryze AI says ChatGPT, Perplexity, Gemini, and Claude now send visitors but GA4 often buries them in Referral or Direct. KISSmetrics goes further: “Most of your ChatGPT traffic is sitting in the direct bucket.”

That should change how you think about attribution.

If you’re posting screenshots of your product showing up in ChatGPT, but you can’t connect that to visits, signups, or revenue, you don’t have proof. You have a nicer-looking vanity metric.

What does attribution mean when AI traffic lands in Direct?

Attribution means knowing which source started a visit or influenced a conversion. With AI traffic, that breaks fast because the source often disappears before GA4 can label it. KISSmetrics says you need to check both the HTTP referrer and utm_source values, then store that signal at the person level because someone might arrive from ChatGPT on Tuesday and convert on Friday. By then, a session-scoped source can be gone.

That matters for small SaaS teams because default GA4 reports can tell the wrong story. If an AI assistant starts demand but the visit ends up in Direct, you can over-credit brand search, email, or word of mouth. If a founder asks, “Is this AI visibility thing working?” GA4 can easily answer with a shrug.

Good attribution is not “we got mentioned.” Good attribution is “this page got cited, this many people clicked, and these visits turned into trials.”

GA4 misses AI traffic for reasons you can’t fix with one report

GA4 is not just messy here. It has a blind spot built into how AI assistants send traffic. Ryze AI recommends custom channel groups, regex filters, Explore reports, and UTM tagging because the default channel setup does not break AI traffic out cleanly. KISSmetrics says part of the problem is structural, not user error.

They point to two practical issues:

• native apps often send no referrer at all
• some assistant surfaces strip referrers but keep UTM tags
• session-scoped attribution gets overwritten before the person converts

KISSmetrics even shares its own one-quarter people-based numbers: ChatGPT 292, Gemini 117, Claude 106, and Perplexity 79. That is not a market benchmark, and they say yours will look different. But it does prove the channel is real enough to bother measuring.

The uncomfortable part is that some of this traffic is still unrecoverable. If a native app sends no referrer and no useful tag, no analytics tool can invent that source after the fact. So the goal is not perfect attribution. The goal is less self-deception.

The pages worth building are the ones you can trace from citation to conversion

AI visibility should change what you publish. It should not push you toward more vague top-of-funnel writing. It should push you toward pages that can earn citations and catch intent when someone actually clicks.

The setup in Ryze AI is a clue. They focus on hostname rules, UTMs, and custom channels because the useful question is not “did an assistant mention us somewhere?” The useful question is “which page got picked up, and what happened after the click?”

For a tiny team, that usually means building fewer pages with clearer jobs:

• pages that answer a specific buying question
• pages with proof, examples, and product detail an assistant can cite
• pages tied to a signup, demo, or trial path you can measure

That is more useful than churning out broad thought-leadership posts and hoping an LLM notices. If one comparison page gets cited and drives trials, you update it. If one opinion post gets mentions but no visits, you stop treating it like a win.

Before you buy another AI content tool, fix your attribution setup

Tooling comes after tracking. KISSmetrics says GA4 will not do this reliably on its own, and Ryze AI only gets clean reporting by adding custom channel logic on top of GA4.

If you run a small SaaS, do these first:

1. group known AI assistant hostnames into their own channel
2. capture assistant UTMs where they exist
3. store first-touch AI source at the person level, not just the session
4. compare AI-sourced visitors against other channels on trials and revenue, not pageviews

That gives you something you can actually use in a weekly review. Not “we showed up in Claude.” Something closer to “these five pages got cited, these two drove trials, and this one wasn’t worth updating.”

That’s the bar now. AI visibility without attribution is just another traffic story founders tell themselves when they don’t have pipeline numbers yet.