How to Track AI Search Referrals in GA4

The practical answer: track AI assistants as a clearly defined referral segment in GA4, preserve the raw source and medium, and report the segment alongside organic search as a separate segment. The result is an observation of visits attributed to known referrers, not a complete count of AI citations or influence.

Start with the question the report can answer

A defensible GA4 report can answer: “How many recorded sessions arrived with a source that matches our declared list of AI assistants?” It cannot tell you how often an answer engine mentioned the brand without a click, whether every assistant preserves referral information, or whether the assistant caused a conversion that happened later through another channel.

Google’s Traffic acquisition report is session-scoped. Its source dimension identifies the publisher or inventory source that sent the session; source/medium combines that source with the acquisition method. Keep both dimensions in the working record so a friendly label never replaces the underlying evidence.

Build a known-referrer register

Create a small versioned table with the assistant name, exact observed source value, dated evidence, and review owner. The former OpenAI publisher FAQ now redirects to Atlas-retirement guidance and no longer supports a current universal referral-tag claim. Retain utm_source=chatgpt.com when observed, then verify each current click route before depending on it.

Minimum fields for an AI-referral register

Field	Record	Why it matters
Assistant	Product name	Creates a readable reporting label
Observed source	Exact GA4 value	Preserves the underlying evidence
Evidence	Official note or dated test	Prevents guessed patterns becoming rules
Checked	Date and owner	Makes maintenance explicit
Boundary	Click referrals only	Stops the metric being presented as citation coverage

Inspect the raw sessions before creating a channel

1. Open Reports → Acquisition → Traffic acquisition.
2. Change the primary dimension to Session source / medium.
3. Search for each documented source value separately.
4. Add landing page, country, device category, and conversion-related metrics only when they help answer a declared decision.
5. Export the rows with the date range, property, filters, time zone, and consent state.

Start with this raw inspection because channel rules can hide mistakes. A source may be absent, uncommon, or grouped in an unexpected place. Zero recorded sessions means zero sessions matching the current query and measurement conditions; it does not prove the assistant never used or cited the site.

Check the default AI Assistant channel before creating a custom group

Google’s current default-channel definitions include an AI Assistant channel. The rule uses the ai-assistant medium and a maintained referrer list. Google’s AI Overviews and AI Mode remain in Organic Search. Inspect Session source / medium alongside Session default channel group before deciding whether a custom group adds useful coverage.

A custom channel can organize signals that exist; it cannot reconstruct a missing referrer. Keep its rules versioned and compare them with the default classification. A fall in identified ChatGPT sessions warrants investigation, but a rise in Direct is not enough to reassign those visits to ChatGPT.

GA4 custom channel groups let editors define rule-based channels using dimensions such as source and medium. Google documents that custom channel groups can be applied retroactively in reports. That makes them useful for a maintained reporting view without rewriting historical event data.

1. In Admin, open Data display → Channel groups.
2. Create a group named with a version, such as AI assistants v1.
3. Add one channel whose source rules match only the entries in the evidence register.
4. Keep organic search, referral, email, paid, and direct as separate channels.
5. Save a screenshot or export of the rules and record every later change.

Use exact matches where possible. If a regular expression is necessary, test it against a list containing both expected matches and dangerous near-matches. A broad expression containing “ai” can capture unrelated domains and corrupt the report.

Build the weekly decision view

A useful weekly view is small: sessions, engaged sessions, key events or conversions, landing pages, and source. Compare a complete current period with the previous comparable period, then inspect whether a change is concentrated in one source or landing page. Annotate releases, tracking changes, consent changes, and campaigns.

Metrics and the claims they can safely support

Metric	Useful for	Does not prove
Sessions	Recorded visit volume	Total AI citations or mentions
Engaged sessions	GA4 engagement under current configuration	Content quality by itself
Landing page	Which destination received the visit	Which passage was cited
Key events	Configured on-site outcomes	Incremental revenue caused by AI search

Validate with a controlled click test

When a product visibly links to your site, run a bounded test in a non-production analytics property or clearly labeled test window. Record the assistant, interface, link, browser, consent choice, destination URL, timestamp, resulting source/medium, and any campaign parameter. Repeat on more than one device or browser when that distinction matters.

The test establishes how that route was recorded in that environment. It does not prove all clicks from the product use the same path. Keep the register open to correction and mark entries as official, observed, or unverified.

Report the limits beside the number

Every dashboard should state that the segment covers recorded click referrals matching the current rule set. It may omit citations without clicks, stripped referrers, privacy-restricted sessions, new products not yet registered, and journeys completed through another channel. This boundary is not a footnote; it is what makes the number usable.

Use a decision rule, not a vanity chart

Name the decision before adding the report to a dashboard. A content team might use recurring qualified visits to identify landing pages worth maintaining. A product team might compare assistant sources to find broken destinations. A founder might monitor whether the channel produces enough observed value to justify a dedicated measurement or publishing workflow. Each decision needs an owner, review frequency, minimum sample, and action.

• Expand: a repeatable source sends qualified visits to pages that satisfy the expected job.
• Repair: visits arrive but land on broken, stale, or mismatched destinations.
• Hold: the sample is too small or the reporting route is not stable enough to interpret.
• Stop: maintaining the segment costs more than the decision value it creates.

Never use a low total as permission to remove ordinary search fundamentals. The same page can be discovered through Web search, direct links, newsletters, social platforms, and answer products. Channel reporting should inform allocation while preserving the page’s full reader job.

Reporting example

A referral report should preserve the source that actually sent the visit

If every unfamiliar hostname is rewritten as “AI traffic,” the report becomes impossible to audit.

1. Capture the original source and medium before channel grouping.
2. Maintain a dated lookup table for known answer-engine referrers.
3. Keep direct and unknown traffic separate instead of forcing attribution.
4. Compare landing pages and outcomes, not sessions alone.

My takeaway: The best GA4 view is a transparent segment that can be revised when referral behavior changes, not a permanent black-box channel.

Primary documentation

• Google Analytics: Traffic acquisition report
• Google Analytics: Create and edit custom channel groups
• OpenAI: crawler reference

When identified ChatGPT visits fall, trace one arrival

Use a consenting test visit and record the visible citation destination, redirect chain, final landing URL, available referrer, consent state, and observed GA4 source/medium. Redact session identifiers before sharing evidence. Keep the client, date, and product surface with the record because a mobile app, desktop browser, and API citation can follow different paths.

If the final URL lost a campaign parameter, inspect the redirects before blaming the assistant. If a tag survives but GA4 differs, inspect collection and classification. If both tag and referrer are absent, report the source as unresolved unless another instrument establishes it. This is a proposed diagnostic procedure; no controlled click test or site-wide attribution-loss estimate was produced for this update.

Correction note: The October 15 update adds the documented default AI Assistant channel and replaces the former FAQ-based claim that every ChatGPT referral carries the same tag.