Whether all the work on AI visibility pays off comes down to one sober question: do visitors arrive through it? The answer is in GA4, but it is not there by itself — AI visitors hide in the default view under "Referral", between trivia. With ten minutes of setup they become visible, permanently and comparable with your other channels.
This piece shows both paths — the quick filter and the clean permanent solution —, names the blind spots no setup in the world removes, and explains how to read the initially small numbers without over- or underrating them.
How GA4 recognizes an AI visitor
When someone clicks your source in a ChatGPT answer, the browser sends the origin along, and GA4 records it as the session source. The names to look for:
chatgpt.com— ChatGPT; OpenAI additionally appendsutm_source=chatgpt.comto links, so these visits stay attributable even when the referrer gets lostperplexity.ai— Perplexitycopilot.microsoft.com— Microsoft Copilotgemini.google.com— the Gemini appclaude.ai— Claude
The list keeps growing, but these five cover nearly everything today. Two quirks up front: visits from AI apps on phones sometimes arrive without an origin and land as direct traffic — the measured number is a lower bound. And the biggest AI channel of all is missing from this list entirely; more on that in a moment.
The quick look: filtering a report
For a first inventory, the standard report is enough. Open Reports → Acquisition → Traffic acquisition, set the primary dimension to "Session source / medium" and add a filter on top: dimension "Session source", match type "matches regex", value:
.*(chatgpt|openai|perplexity|claude|copilot|gemini).*
The result is your AI traffic list: which systems, how many sessions, and via the "Landing page" dimension also which of your pages get cited. If almost nothing shows up today, that is not a verdict against the work — it is the baseline you compare against in six months.
The permanent solution: a custom channel group
To make AI traffic appear as its own channel next to Organic Search and Direct in every report, create a custom channel group: Admin → Data display → Channel groups, copy the default group, and create a new channel "AI" inside it. As the condition, use the same logic as the filter: source matches the regex .*(chatgpt|openai|perplexity|claude|copilot|gemini).*.
Two details decide the data quality. First, the order: drag the AI channel above "Referral" and "Organic Search" in the channel list, because GA4 sorts every session into the first channel whose condition matches — if "Referral" comes first, it swallows the AI visits as before. Second, the view: the new group is an additional dimension (swap "Default channel group" for yours), reports do not switch to it on their own.
The reward: every report that breaks down by channel now carries AI as its own row — with the same metrics as any other channel, from engagement rate to conversions. How to read those reports is covered in Reading GA4.
The blind spots
Three things no GA4 setup will show you, and it is better to know them than to hunt for them in the numbers.
Clicks from Google's AI Overviews look like normal Google clicks. The overview is part of the results page, so the click arrives as google / organic — indistinguishable from a click on a blue link. The AI channel with the biggest reach is thus invisible in GA4 by design; what can still be read indirectly is covered in Google's AI Overviews.
The answer without the click. Most AI usage produces no visit at all: the question gets answered, the user is satisfied, your website played its part as a source and nobody clicked. That visibility still works — as brand contact and as a recommendation — but it leaves no trace in GA4. GA4 measures the tip of the iceberg.
Lost origins. Whoever copies a link from an AI answer and opens it later, or arrives from an app without a referrer, counts as direct. An unexplained rise in direct traffic alongside growing AI visibility is therefore often no coincidence.
How to read the numbers
The most common mistake is judging AI traffic by its volume. Session counts are small on almost every website today — single-digit percentages would already be a lot. Two other readings are the interesting ones.
First, quality. Compare the AI channel against your average: engagement rate, engagement time, and above all conversions — set up cleanly as described in Setting up conversion tracking. Visitors from AI answers have already had their question answered and click anyway; these are often the most pre-qualified visitors a website can get. If this channel converts better than the rest, every further investment in quotable content is justified — with your own data instead of industry prose.
Second, direction. Chart the AI sessions monthly and look at the slope, not the level. A channel that doubles every few months is strategically more important than its current share suggests — and exactly this pattern is what AI traffic currently shows on many websites.
And third, as a bonus: the "Landing page" dimension inside the AI channel shows which of your pages the answer systems cite. That is your most direct feedback on which content works in the sense of Generative Engine Optimization — and which topics deserve a second page like it. It doubles as the shortlist for your llms.txt: the pages answer systems already reach for are the ones a map of your site should name first.
Frequently asked questions
How do I see in GA4 whether visitors come from ChatGPT?
As the session source chatgpt.com — OpenAI additionally tags links with utm_source=chatgpt.com. The fastest way is a filter on the acquisition report with a regex across the known AI sources; the cleaner permanent way is a custom channel group with an AI channel.
Why does traffic from Google's AI Overviews not show up?
Because the overview is part of the normal Google results page: a click on a cited source arrives as google / organic and cannot be told apart from an ordinary search click. Neither GA4 nor Search Console reports overviews separately.
How much AI traffic is normal?
On most websites today, under one to two percent of sessions — what matters is not the share but the growth and the quality. AI visitors often convert above average, because they arrive with their question already answered and a recommendation at their back.
Does a custom channel group work retroactively?
Yes. Channel groups are computed when reports are queried, not when data is collected — the new AI channel therefore displays past periods correctly too. What matters is only the channel's position above "Referral", so its condition fires first.
Why does my AI traffic fluctuate so much?
Because the answer systems' source selection is not deterministic: the same question can cite your page today and another one tomorrow. With small absolute numbers this comes through as heavy noise. Judge the channel over months, not days.