What referrers indicate AI traffic?

The main ones are chatgpt.com and chat.openai.com for ChatGPT, perplexity.ai and www.perplexity.ai for Perplexity, claude.ai for Claude, copilot.microsoft.com and bing.com/chat for Microsoft Copilot, and gemini.google.com for Gemini. Google AI Overviews are the exception: clicks from them arrive as ordinary google.com organic search referrals and cannot be separated in standard analytics.

Why is my AI referral traffic so low?

Three reasons compound. Answer engines are designed to resolve the question in-place, so most useful answers never produce a click at all. Some clients open links in a way that drops the referrer, which pushes visits into your direct bucket. And AI Overviews clicks are indistinguishable from organic Google. The visible number is a floor, not a measurement of influence.

Is AI referral traffic worth optimising for?

Optimise for citation, and treat referrals as one imperfect proxy for it. Referral sessions from answer engines typically show higher intent than organic search — the user has already been told what you do and clicked anyway — but the volume is too small and too undercounted to be a primary KPI. Pair it with crawler-side data on which of your pages answer engines actually retrieve.

Blog Measurement

Tracking AI referral traffic in GA4 and beyond

The referrers to segment on, why the numbers look small, why they undercount, and how to connect a referral back to the page the model actually cited.

July 25, 2026 · Updated August 7, 2026 Markdown for agents

There are two halves to AI visibility. The crawler side — who fetches your content — needs server-side measurement, because crawlers never run your analytics tag. The referral side is the easier half: when a person clicks a citation in ChatGPT or Perplexity, a real browser loads your page and your existing analytics sees it. You just have to build the segment, and understand what the resulting number does and does not mean.

The referrers to match

Source	Hostnames
ChatGPT	`chatgpt.com`, `chat.openai.com`
Perplexity	`perplexity.ai`, `www.perplexity.ai`
Claude	`claude.ai`
Microsoft Copilot	`copilot.microsoft.com`, `bing.com/chat`
Gemini	`gemini.google.com`
Meta AI	`meta.ai`
Poe	`poe.com`
You.com	`you.com`

Google AI Overviews is the significant omission. Clicks from an AI Overview carry a google.com referrer with organic search parameters and are not separable in GA4 or in Search Console. If AI Overviews are a large part of your category’s surface, accept that the effect will show up as changes in organic click-through rate rather than as a distinct source.

Setting it up in GA4

GA4 will already record these as referral traffic. The work is grouping them so they roll up as one channel instead of eight unrelated rows.

Create a custom channel group under Admin → Data display → Channel groups. Add a channel named AI assistants, placed above Referral in the ordering, with the condition Source matches regex:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|copilot\.microsoft\.com|gemini\.google\.com|meta\.ai|poe\.com|you\.com

Channel groups apply going forward and, in GA4, retroactively to reporting for data already collected — but the ordering matters, since a session is assigned to the first matching channel.

For anything more than counting, add a landing-page dimension to the report. Which page a model cited is far more useful than the total, because it tells you which content is doing the work.

If you use something else

Plausible, Fathom, Umami, and PostHog all expose the referrer directly and support a filter or saved segment on the same hostname list. In PostHog it is a cohort or an insight filter on $referring_domain; in Plausible a saved segment on the Sources report.

Why the numbers undercount

Answer engines are built to prevent the click. A good answer resolves the question in the interface. Citations are attribution, not navigation. Your content can shape thousands of answers and produce a few dozen sessions, and that is the system working as designed, not a failure of your content.

Referrers get dropped. Some clients open links in contexts that strip the Referer header — native app webviews and certain privacy configurations. Those visits land in direct, mixed with bookmarks and typed URLs, unrecoverable.

Not every mention is a link. A model that says “Shiftrank does AEO observability” without a citation generates no traffic and no measurable event, but it is doing exactly what brand marketing is supposed to do.

Treat the referral number as a floor and a trend line. Do not present it to a board as the measurement of AI performance.

Connecting referrals to citations

The interesting question is not how many AI sessions you had. It is which page the model chose to cite, and why that page.

You can approximate the answer by joining two datasets:

Crawler-side retrieval. Which URLs the live fetchers — ChatGPT-User, Perplexity-User, Claude-User — request, and how often. A page repeatedly fetched by live agents is a page the model keeps deciding is relevant to real questions.
Referral landing pages. Which URLs actually receive the resulting clicks.

Pages that appear in the first list but not the second are being read and used without earning a click — that is influence without traffic, and it is usually fine. Pages in both are your citation earners; find what they have in common and do more of it. Pages in neither, despite being commercially important, are the gap worth working on.

The first dataset requires server-side collection, since none of those fetchers execute JavaScript. Measuring AI crawler traffic covers the four ways to get it. Shiftrank records both halves in one place, because the join is the point.

What to report

A monthly view that has held up:

Retrieval coverage — distinct URLs fetched by live AI agents, and whether your priority pages are among them
Top retrieved paths — the demand signal, ranked
AI referral sessions — by assistant, with landing page
Conversion rate of AI sessions against organic search, so the small volume gets its intent premium acknowledged
Coverage gaps — priority pages that AI agents are not fetching, as an explicit to-do list

That last row is the one that changes behaviour. Everything above it is context.

Common questions

What referrers indicate AI traffic?: The main ones are chatgpt.com and chat.openai.com for ChatGPT, perplexity.ai and www.perplexity.ai for Perplexity, claude.ai for Claude, copilot.microsoft.com and bing.com/chat for Microsoft Copilot, and gemini.google.com for Gemini. Google AI Overviews are the exception: clicks from them arrive as ordinary google.com organic search referrals and cannot be separated in standard analytics.
Why is my AI referral traffic so low?: Three reasons compound. Answer engines are designed to resolve the question in-place, so most useful answers never produce a click at all. Some clients open links in a way that drops the referrer, which pushes visits into your direct bucket. And AI Overviews clicks are indistinguishable from organic Google. The visible number is a floor, not a measurement of influence.
Is AI referral traffic worth optimising for?: Optimise for citation, and treat referrals as one imperfect proxy for it. Referral sessions from answer engines typically show higher intent than organic search — the user has already been told what you do and clicked anyway — but the volume is too small and too undercounted to be a primary KPI. Pair it with crawler-side data on which of your pages answer engines actually retrieve.

Keep reading

How to measure AI crawler traffic (and why your analytics shows zero)
Google Analytics cannot see GPTBot. Here are the four ways to capture AI crawler traffic, what each one costs you, and the five metrics worth reporting once you have the data.

See which answer engines already read this page

Shiftrank records every AI crawler hit from the moment your DNS resolves.

Start free