Start with what the browser actually exposes
AI referral measurement begins when a person clicks from an assistant or generated-answer surface to the website and the browser preserves a source signal. That signal may be a document referrer, a tagged URL or both. Sometimes it is absent, in which case the session should remain direct or unknown according to the analytics system's rules.
No analytics setup can reconstruct every journey. Mobile apps, privacy controls, redirects, copied links and cross-device behavior can remove the referrer. The objective is defensible first-party measurement, not perfect surveillance.
A four-layer measurement model
| Layer | Purpose | Typical fields |
|---|---|---|
| Request | Separate known bots and fetchers from human navigation | timestamp, path, user_agent, bot_class, response_status |
| Session | Preserve the first observable acquisition context | session_id, landing_page, referrer, source, medium, started_at |
| Event | Record meaningful on-site actions | event_name, page, timestamp, session_id |
| Lead | Connect an enquiry to consented session context | lead_id, session_id, submitted_at, qualification, attribution fields |
Build an explicit referrer taxonomy
Use hostnames rather than loose text matching. A maintained list may include currently observed destinations such as chatgpt.com, perplexity.ai or copilot.microsoft.com. Keep the list versioned and review new values manually before assigning them to the AI-referral channel.
- Store the raw referrer before classification.
- Normalize hostname case and remove irrelevant path noise.
- Keep source and medium separate.
- Do not classify every bing.com referral as Copilot; the hostname can serve several products.
- Preserve an “unknown / needs review” bucket.
- Record the taxonomy version used in each report.
If a platform or campaign gives you control over a destination URL, use consistent UTM parameters. Do not add invented UTMs after the visit or use them to overwrite a genuine external referrer.
Implementation sequence
- Capture acquisition once. On the first page view, store the raw referrer, landing URL, UTMs and timestamp in the consented session context.
- Classify after capture. Map the hostname to a documented source category without deleting the raw value.
- Exclude known bots. Use server or edge data where available; do not rely on engagement time alone.
- Send analytics fields. Preserve source, medium, landing page and campaign values in GA4 or the chosen analytics system.
- Connect the form. Pass a session identifier or a small approved attribution payload to the lead record.
- Qualify separately. A form submission becomes a qualified lead only after the business applies its qualification rule.
A minimal lead attribution payload
Fields worth preserving
first_known_source, first_known_medium, first_landing_page, last_known_source, session_started_at, lead_submitted_at and attribution_confidence.
Do not place confidential prompt text, account identifiers or browsing history in the lead payload. Collect only what the business has a legitimate, disclosed reason to use.
A validation protocol before reporting
- Open a controlled tagged link and confirm the landing URL retains its parameters.
- Run a controlled visit with an observable referrer when the platform allows it.
- Confirm one human session appears with the expected source and medium.
- Submit a clearly marked test form and verify that the session context reaches the test lead.
- Check server logs to ensure the test was not classified as a crawler.
- Delete or exclude test records from production reporting.
What to report monthly
| Metric | Report with |
|---|---|
| AI-referral sessions | Raw count, source hostnames, landing pages and bot exclusions |
| Engaged sessions | Definition used and analytics configuration |
| Enquiries | Count by observable first and last known source |
| Qualified leads | Qualification rule and review owner |
| Unknown/direct | Count retained rather than reassigned |
Privacy and measurement boundaries
Consent, retention and access rules should be decided before connecting session data to a lead. Avoid collecting prompt content or sensitive identifiers. Referrer data can support attribution, but it does not reveal the full decision path and should not be presented as deterministic proof.
Primary analytics references
- Google Analytics: Campaigns and traffic sources
- Google Analytics: Traffic-source dimensions and manual tagging
- Google Analytics: Direct / none traffic
Where to go next
Use the Nexus AI Visibility Framework to place this topic inside a complete measurement system, or review the AI Visibility Audit scope.
13 Jul 2026 — Expanded with an independent structure, examples, implementation guidance and primary references.