How to track AI referral traffic in GA4

To track referral traffic from ChatGPT, Claude, Gemini and Perplexity, preserve the raw page referrer and UTM values, create a maintainable AI-assistant channel grouping in GA4, validate it with controlled visits, and report sessions, engagement and conversions by landing page. Expect undercounting: apps, privacy controls, redirects and copied links can remove referral information. Classify only recognized evidence as AI referral traffic and leave direct or unknown traffic unassigned unless another reliable signal supports the connection.

Build a governed source-domain dictionary

Start with source and referral values actually appearing in your property, then map verified domains associated with ChatGPT, Claude, Gemini and Perplexity. Store the dictionary outside one analyst's memory, include the date each rule was verified, and review it as products and domains change. Match carefully enough to avoid unrelated subdomains or copycat sites. Preserve raw source, medium, referrer and landing URL fields so a bad classification can be corrected later without losing the original evidence.

• Recognized referral: a verified assistant-related referrer reaches the site.
• Tagged referral: a controlled link carries a documented campaign parameter.
• Self-reported AI: the visitor or buyer names an assistant in a neutral survey.
• Possible AI: behavioral or direct-traffic pattern without source-level proof.
• Unknown: no dependable signal; retain it rather than reallocating it.
• Excluded: internal traffic, testing, bots, payment redirects and known misclassifications.

Configure GA4 without rewriting history

Create an exploration or custom channel definition that groups known assistant sources, and keep the original default-channel view available. Add landing page, source/medium, device, geography and new-versus-returning dimensions. Mark business events only after verifying that they fire once and carry the intended parameters. Channel-rule changes should be annotated with an effective date; avoid presenting a newly expanded domain list as organic traffic growth. If your consent setup limits analytics, disclose that coverage instead of extrapolating silently.

Validate sessions and diagnose missing referrers

Use controlled, permissioned test links from each available surface and inspect real-time or debug data. Confirm the landing page loads, query parameters survive redirects, cross-domain settings behave correctly and self-referrals are excluded. Then compare browser requests, server logs and GA4 collection at an aggregate level appropriate to your privacy obligations. Some assistant clicks will arrive as direct or have no usable referrer. Do not fingerprint visitors or weaken consent controls to recover an attribution label.

1. Save existing channel definitions and raw source values before adding new rules.
2. Create a verified domain and campaign mapping with an accountable owner.
3. Run controlled visits and inspect source, medium, landing page and event collection.
4. Build a report for sessions, engaged sessions, key events and conversions.
5. Compare assistant sources and landing pages without assuming equal user intent.
6. Review rules quarterly and annotate every material classification change.

Connect referral behavior to answer evidence

Referral traffic answers "who arrived with identifiable source evidence," while AI visibility monitoring answers "what sampled assistants said." Connect them by date, topic and landing page for investigation, not by claiming a monitored answer generated a particular session. Use the AI visibility scan for answer evidence and the AI readiness checker for landing-page access and clarity. Monitor and higher plans include trend monitoring and Google integrations; Monitor supports one connection for either GSC or GA4, so verify current plan limits. Report referral sessions, assisted conversions and pipeline separately from mentions, citations, recommendations, sentiment, accuracy and share of voice.

Evaluate the landing-page experience, not only channel volume

Review landing pages receiving identifiable AI traffic as a distinct audience segment, but avoid personalizing from sensitive inferences. Ask whether the page answers the likely question immediately, supports its claims and offers a proportionate next step. Compare engagement with similar landing pages while controlling for device, geography and intent where possible. A low conversion rate may indicate an informational visit rather than poor traffic. Pair aggregate behavior with voluntary survey responses and usability research before changing the experience. Watch for a small number of unusual sessions distorting averages, and use medians or distributions where appropriate. Document which page changes follow the analysis, then compare a matched period without assuming the channel caused every difference. The tracking program succeeds when classifications are reproducible and decisions improve, not when the AI channel looks artificially large.