How to Track Traffic from ChatGPT, Perplexity, Gemini & AI Overviews

“Traffic from AI” is actually three distinct things, and conflating them is the most common mistake in AI-era reporting:

1. 1AI referral traffic — a human read an AI answer, clicked a cited link, and landed on your site. This is real, measurable, human traffic — if your analytics recognizes the referrer.
2. 2AI crawler traffic — bots (GPTBot, ClaudeBot, PerplexityBot…) fetching your pages for training or for real-time answering. Not human, invisible to JavaScript analytics, visible only in server logs and CDN dashboards.
3. 3AI-influenced traffic — a human was told about you by an AI and arrived later via branded search or a typed-in URL. The largest layer by most estimates, and structurally invisible to click tracking. Measurable only by proxy.

This guide covers how to capture layer 1 precisely, monitor layer 3 honestly, and read layer 2 for what it is: a leading indicator, not an audience. For where AI-influenced journeys break attribution models themselves, see our attribution models comparison.

Each assistant handles outbound links differently, which is why AI traffic looks inconsistent in analytics. Here is the fingerprint each major platform leaves (behavior as of mid-2026 — platforms change this without notice, so re-verify quarterly):

The largest AI answer surface on the internet is also the one you cannot isolate. Clicks from Google's AI Overviews and AI Mode arrive with a standard google.com referrer, so they land in Organic Search alongside classic blue-link clicks. Google Search Console counts AI Overview impressions and clicks inside regular search totals and offers no segment for them.

What you can observe is the aggregate symptom: since AI Overviews rolled out broadly, independent studies have repeatedly measured significantly lower click-through rates on informational queries where an Overview appears. In practice, many sites see impressions holding steady or rising while clicks decline — a widening impressions-to-clicks gap is often the fingerprint of AI Overviews absorbing your former traffic.

• Monitor the ratio, not just the totals: in Search Console, watch CTR trends on your top informational queries specifically — those are Overview-exposed. Falling CTR at stable position is the tell.
• Segment branded vs. non-branded queries: AI-influenced buyers often return via branded search, so branded impression growth alongside non-branded CTR decline is a consistent AI-era pattern.
• Accept the limit: any tool claiming to give you exact AI Overview referral counts from analytics data is modeling, not measuring. Visibility inside the Overviews themselves is measurable — by monitoring the answers, not the clicks (see AI visibility tools compared).

GA4 has no native “AI” channel, but custom channel groups solve it cleanly. The goal: every session whose source matches a known AI platform gets classified into an “AI” channel you can trend, segment, and report on.

1. 1In GA4: Admin → Data settings → Channel groups. Don't edit the default group — create a copy (e.g. “Channels incl. AI”) so you keep a clean baseline.
2. 2Add a new channel, name it “AI Assistants”, and drag it above Referral and Organic Search in the evaluation order — channel groups apply the first matching rule.
3. 3Set the condition: Source matches regex with a pattern like the one below.

.*chatgpt\.com.*|.*chat\.openai\.com.*|.*perplexity\.ai.*|
.*gemini\.google\.com.*|.*copilot\.microsoft\.com.*|
.*claude\.ai.*|.*grok\.com.*|.*x\.ai.*|.*meta\.ai.*|
.*deepseek\.com.*|.*edgeservices\.bing\.com.*|.*mistral\.ai.*

(Written as one alternation, on one line, in the GA4 field. GA4 regex here is a full match, hence the .* wrappers; escape the dots as shown. Because ChatGPT appends utm_source=chatgpt.com, matching on source catches both its referrer and its UTM variants.)

• Caveats: custom channel groups are not retroactive in some reporting surfaces, so build this now even if AI volume is small — you're creating your future baseline.
• For historical analysis, an Exploration filtering session source on the same regex works on past data.
• Re-check the regex quarterly. New assistants appear, domains change (chat.openai.com → chatgpt.com is the canonical example).
• Expect the volumes to be small — commonly low single-digit percentages of sessions — but watch two things: the growth rate and the conversion rate. Many teams find AI-referred visitors convert meaningfully better than average, which makes sense: they arrive pre-researched, often having already compared you with alternatives inside the assistant.

Even a perfect channel setup undercounts AI referrals. The main leaks:

• Native apps: clicks from AI mobile and desktop apps frequently pass no referrer and land in Direct. Given how much assistant usage is app-based, this is a structural undercount, not an edge case.
• Copy-paste behavior: assistants often present URLs users copy into a browser rather than click. Direct again.
• Referrer stripping: privacy settings and browser policies can strip cross-origin referrers.
• Consent declines: in consent-strict markets, a slice of all traffic — AI included — never reaches GA4 at all.

The honest posture for client reporting: treat measured AI referrals as a floor, never the total. An unexplained rise in Direct traffic that coincides with rising AI referrals and rising branded search is usually the same story told three ways.

Before an assistant can cite you, it has to read you. AI crawler activity is measurable in server logs or CDN analytics (Cloudflare and others now ship AI-crawler dashboards) by user agent:

Two practical uses. First, on-demand fetchers (ChatGPT-User, Perplexity-User, Claude-User) are a real-time signal that assistants are consulting your pages to answer live user questions — trending which URLs they hit tells you which content is doing invisible saleswork. Second, robots.txt policy is now a measurement decision: blocking AI crawlers may be a defensible content-rights position, but understand that it also removes you from consideration in a growing research channel. That trade-off belongs in client strategy conversations, not in a default robots.txt template.

Most AI influence never becomes a click, so the measurement stack needs proxies. The four that work, in rough order of practicality:

1. 1Self-reported attribution: add “an AI assistant (ChatGPT, Perplexity…)” as an explicit option in your “how did you hear about us?” field. The gap between this number and your AI-referral sessions is your dark-AI multiplier — and it's often large.
2. 2Branded search volume: assistants create demand that resolves via Google. Trend branded impressions in Search Console against your AI-visibility efforts.
3. 3Direct-traffic trends segmented to new visitors on desktop (where app-based, referrer-less AI clicks concentrate).
4. 4Share of voice in AI answers: monitor the answers themselves — how often, and how favorably, assistants mention a brand across a prompt set. This is upstream of all click data; we compare the tooling in AI visibility monitoring tools compared.