Track AI referral traffic in GA4: the AI Assistant channel

Where did AI visits appear before the dedicated channel?

Earlier assistant-origin visits did not disappear simply because GA4 lacked one convenient channel label. A ChatGPT link might contain a source parameter such as ?utm_source=chatgpt.com, or the browser could send referral information. GA4 might identify the source while grouping the session with other referrals. If no usable source information arrived, a visit could appear as Direct.

Knowing the source and grouping similar sources are separate operations. Analysts previously built their own rules to bring ChatGPT, Gemini or Claude visits together. The dedicated channel reduces that classification work for visits matching Google’s definitions. It does not recover missing source information or ensure that every assistant-origin visit is identifiable.

That distinction also matters when comparing periods around the introduction of a channel. A classification change is not, by itself, evidence that a newly labelled visit represents new demand. Keep the date range and source definitions visible in your report.

Separate crawler access, answer citations and human visits

For example, GPTBot or ClaudeBot retrieving an article is a request by software. A person clicking a restaurant’s official website from an assistant’s answer is a different event. If the person reads the answer and leaves without clicking, the citation creates no corresponding website session in GA4.

In this guide, AI referral traffic means identifiable visits after someone follows an AI link. For crawler policy, use the separate robots.txt guide. For mentions and citations without clicks, use answer observations such as those discussed in the AI search retrieval study.

Growth is a reason to inspect your data, not assume superior visitors

Similarweb’s September 2026 analysis estimated an average of about 770 million AI referral visits per month worldwide from June 2025 through May 2026, up 117.4% year on year. This is a desktop and mobile-web estimate, not a census of measured sessions from every website. Rapid growth can coexist with a small share of total traffic.

The Adobe retail results and the Kaiser–Schulze study differ in time period, platforms, sample composition and comparisons. One does not automatically refute the other. Neither establishes that AI visitors to your particular site will behave better than other visitors.

That is why landing pages and observed actions matter more than importing an external percentage into a forecast. A technical article and a product checkout page have different jobs; the visitors who reach them may have different intentions before the first page view.

How GA4 assigns AI Assistant

Google’s default channel definitions classify AI Assistant when the medium exactly matches ai-assistant. When the referrer matches its recognised list of assistants, the medium is set to ai-assistant and campaign to (ai-assistant). These are reporting values, not proof that you bought an advertising campaign.

The channel includes recognised assistant sources such as ChatGPT and Gemini. Google AI Overviews and AI Mode are explicitly excluded and remain part of Organic Search. Do not add their supposed traffic to AI Assistant merely because the interface includes generated text.

Source information is not always available. Opening a link from an app, redirects or tracking restrictions can interrupt the trail. If no other usable campaign or source information is present, Direct may contain some of those visits. Direct also contains many other visits, so assigning its total to AI would invent attribution.

GA4 distinguishes session-scoped, first-user and event-scoped channel dimensions. The workflow below uses Session default channel group consistently to examine visits. Switching scope part way through can make two superficially similar reports answer different questions.

A four-step reporting workflow

Choose one date range before starting and keep it through the comparison. Report names and available collections can differ by property configuration; the important parts are the session-scoped dimensions and the retained filter.

1. Open Traffic acquisition

Go to Reports → Acquisition → Traffic acquisition. Confirm the primary dimension is Session default channel group. AI Assistant belongs alongside channels such as Organic Search, Direct and Referral.

2. Filter the channel

Add a filter for Session default channel group exactly matching AI Assistant. Read the sessions count as visits, not people: one person may start multiple sessions. An illustrative total of 100 sessions therefore does not mean 100 distinct individuals. Retain the channel filter when changing the report dimension.

3. Connect each source to its landing pages

Switch to Session source / medium to separate identifiable assistant sources. Then open the Landing page report, apply the AI Assistant filter there, and add Session source / medium as a secondary dimension. A landing page is the first page of that session, rather than every page the visitor reads.

This connects the source with the content or product page receiving the visit. Keep the actual URL visible when different pages have similar titles, and compare the same period. Do not assume that a source total in one report corresponds to an unfiltered landing-page total in another.

4. Examine behaviour and important actions

Return to Session default channel group, remove the AI Assistant-only filter and compare with appropriate channels. Keep dates and action definitions aligned. Show counts alongside rates: a single additional purchase can substantially change a rate based on only a few sessions.

If the report is empty or the source looks wrong

An empty AI Assistant row means the selected report contains no matching recorded data under those settings. It does not establish that no one saw your brand in an AI answer.

1. Check the date range, dimension and filter. Verify session scope and the exact channel value before interpreting an absence.
2. Inspect available source / medium values. A recognisable assistant source may need investigation even when a channel comparison looks unexpected. Compare the recorded values with the documented rule rather than guessing from a page title.
3. Check the measurement path. Confirm the relevant page is tagged and investigate redirects or app handoffs that might remove source information. A controlled click can test that path; it cannot recreate every user’s context.
4. Validate the action events. If sessions appear but purchases or sign-ups do not, first check event collection and the selected event definitions before blaming the channel.

Keep unresolved source loss unresolved. Do not rewrite all Direct or unattributed orders as AI-assisted, and do not infer a crawler volume from a GA4 referral report.

Use the report to choose a concrete page improvement

Start with pages already receiving identifiable AI visits. Check that their prices, dates and instructions remain correct, that the answer is easy to find, and that the next step suits the page’s purpose. Updating an outdated tool comparison with recurring ChatGPT referrals has a clearer evidential basis than producing more similar articles without knowing what is wrong.

Compare like with like. One assistant may send more visitors to explanatory articles while another sends fewer directly to product pages. Different purchase rates could reflect landing-page mix, prior intent or a small sample. GA4 cannot show the questions visitors asked before arriving, so those behavioural differences do not establish their cause.

If visits are substantial but the expected action is rare, validate measurement, then inspect whether the page answers the question and makes the next step clear. On an informational page, finding an answer and leaving can be a reasonable outcome. Define the page’s job before judging it.

For recurring review, the relevant output is a short source-and-page report with the period, sessions, engagement, key-event counts, data limitations and a specific follow-up. SEO Pulse is described separately; its availability and output language must be checked rather than assumed from the existence of English documentation.

The report starts at the website visit

AI Assistant makes recognised website visits easier to inspect. It does not reveal all the answers someone read, the alternatives compared or the conversation that influenced a decision. Answer visibility, crawler access, referral sessions and orders remain separate evidence.

Brand mentions, source citations and factual accuracy require observations beyond visit data. The MCAC checks in the GEO guide explain how to keep those outcomes separate.

You can improve a page without measuring every AI exposure first: identify the visits you can observe, keep the content useful and inspect the actions that follow. For a broader measurement discussion, read AI visibility measurement.

Sources and scope

• Google Analytics release notes, 13 May 2026 AI Assistant entry.
• Default channel descriptions and definitions, including AI Assistant and Organic Search.
• OpenAI publishers and developers FAQ.
• Similarweb AI referral traffic by industry, September 2026.
• Adobe’s June 2026 travel and retail report; this guide uses the retail subsection.
• Kaiser and Schulze’s study companion, 973 websites and August 2024–July 2025 data.

No claim here is based on a new LS-Labs analytics experiment. Examples and troubleshooting steps are editorial guidance; the cited external studies retain their own populations and periods.