AI referral source directory: assistants and analytics classification

Major AI referral and search surfaces

The reporting treatment describes the safest current class, while the observed source must still be validated on each property.

Google Analytics documents ChatGPT, Gemini, DeepSeek, Copilot, and Grok as examples of AI Assistants. Other assistants can appear as AI Assistants when they match Google’s maintained list or as Referral under the arriving source. Google AI Overviews and AI Mode remain Organic Search.

Provider apps and browser handoffs can change. Preserve raw approved source evidence, normalization version, and the verification date instead of relying only on a copied domain regex.

Store a durable normalization record

The source name alone is not enough to reproduce a historical classification.

For every mapping, store provider, surface, observed host or source pattern, expected medium, channel class, confidence, rule version, first-seen date, last-verified date, and the primary documentation used. Retain the approved raw acquisition value so a later rule can be audited.

Keep normalization append-only where practical. If a provider adds a domain, publish a new version and effective date rather than silently rewriting historical channel totals.

• Provider and surface
• Observed source pattern
• Expected medium
• Channel class
• Confidence and rule version
• First seen / last verified

Implement the taxonomy without double counting

Use the maintained GA4 default as a reference and add custom logic only for a documented reason.

Compare the property’s Session default channel group with Session source / medium. If a validated provider remains Referral, decide whether a custom primary or custom channel group improves reporting. Keep the default available for comparison.

Apply mutually exclusive rules in a defined order. Exclude internal traffic and machine identities. Test the same session across default, custom, warehouse, and dashboard outputs before publishing.

Recommended order
1. Known paid or campaign rule
2. Validated AI assistant source
3. Google Organic Search
4. Other referral / social / shopping
5. Direct only when no usable source exists

Validate providers with dated controlled visits

A maintained directory needs observations as well as documentation.

For each priority provider, test a public link in the relevant web or app client when practical. Record client, device, destination, source, medium, channel, redirect behavior, timestamp, and result. Do not publish personal URLs or sensitive request data.

Repeat after major provider, browser, redirect, consent, or analytics changes. Mark entries unverified when a result cannot be reproduced and invite providers or practitioners to submit a primary source and dated sample.

Understand what the directory cannot recover

A source taxonomy organizes evidence; it does not create missing evidence.

Copied links, embedded browsers, referrer restrictions, redirects, and cross-device journeys can produce Direct or another later source. The directory cannot identify those sessions by their landing page or content topic.

Likewise, provider naming does not convert crawler requests into human sessions. Use the separate machine directory for user-agent classification and keep answer citations in the visibility stream.

Evidence note

Use observed referral totals as a lower-bound measure of measurable handoffs, not a complete estimate of AI influence.

Change log and correction policy

Versioned maintenance is what makes a reference worth citing.

Version 2.0 expands the directory beyond the largest five examples, adds explicit machine boundaries, and defines a controlled validation record. The downloadable CSV carries the same verification date and should change with the visible table.

Submit corrections to hello@wandered.io with the provider or analytics documentation, observed value, client context, and date. Entries are reviewed for evidence quality before publication; previous versions remain part of the public record.

• Primary source required
• Dated observed example encouraged
• No secrets or personal request data
• Visible table and download updated together
• Prior version preserved
• Material changes described