Measurement Framework Guide
AI referral attribution: known sources and missing data
Read GA4 channel definitions, preserve source evidence through your CRM and compare self-reported discovery separately. Missing referral data cannot identify AI traffic on its own.
v1.1 · October 2026 · Tenten GEO Research
CH.01
Unknown source does not identify the missing channel
A visit with incomplete source information may have come through many routes. Treat the missing information as a measurement limit; do not assign it to AI simply because an assistant could have been involved.
Averi’s April 14, 2026 article reports roughly 57% direct/unknown for its own site. That is one publisher’s example, not an industry rate or the proportion caused by AI. Its summary names GA4 while its body describes a 30-day Fathom analysis, so the measurement basis is not fully consistent. Email, messaging, copied links and typed addresses can also lack usable source evidence.
Chapter takeaway
Keep unknown traffic visible as unknown. A source claim needs evidence beyond an empty referrer.
CH.02
Read current GA4 channel definitions
GA4 now defines an AI Assistant channel for recognized assistant referrers. Google AI Overviews and AI Mode remain within Organic Search. A channel is a reporting rule, not a complete record of everything that influenced a visitor.
Direct and Unassigned have different definitions. Direct uses source (direct) with medium (none) or (not set); Unassigned means no channel rule matched. Also distinguish session, first-user and event-scoped channel dimensions before comparing results. Do not combine them into a single “direct/unknown” metric or treat all reported AI traffic as a guaranteed lower bound.
| Observation | Interpretation | Keep distinct | Next check |
|---|---|---|---|
| Recognized assistant referrer | May be classified as AI Assistant | Attribution scope and actual source | Read current channel rules and the recorded values |
| Google AI Overview or AI Mode visit | Included in Organic Search | Search surface versus channel | Do not relabel all Google organic visits as AI |
| Direct | Matches the specific source/medium rule | Unknown influence versus known entry data | Inspect available landing and campaign evidence |
| Unassigned | No channel definition matched | A rule mismatch versus a missing referrer | Inspect source, medium and tagging configuration |
Based on Google’s current channel definitions; classification does not prove a visitor’s intent.
Chapter takeaway
Report the chosen scope and raw source values. Channel totals, missing records and Google Search traffic answer different questions.
CH.03
Server-side tagging processes evidence it receives
A server container can validate and route measurement events. It cannot reconstruct a missing origin from a generic browser user agent or landing-page path. If the request contains no reliable source evidence, keep the source unknown.
Google distinguishes the event’s page_referrer from the HTTP Referrer header received by the server container. The latter may describe the measurement request rather than the visitor’s original journey. Preserve what was actually received, along with collection time and rule version. A first-party label is still a recorded interpretation; it is neither unerasable nor independent proof.
| Evidence | Useful handling | Limit |
|---|---|---|
| Recorded page referrer | Preserve and validate the actual value | May be absent or incomplete |
| Controlled campaign parameters | Document where the link was distributed | A reused link may travel through other channels |
| Landing path and generic user agent | Use for technical investigation | Do not identify an AI assistant by themselves |
| Consent and storage state | Pass the relevant state through the collection path | Server-side processing does not bypass consent or retention limits |
These are evidence-handling boundaries, not a method for recovering every missing source.
Chapter takeaway
Improve the handling of available data before inventing a source for missing data. Server-side and client-side implementations both need explicit rules and validation.
CH.04
Carry source evidence into CRM without turning it into a score
Keep the technical source, reported discovery channel and qualification outcome in separate fields. AI referral alone does not establish buying intent or justify a higher lead score. Evaluate that relationship using your own qualified outcomes.
Conductor’s 2026 report describes U.S. monthly data from May–September 2025 across 1,215 enterprise customer domains in ten industries. It reports AI referrals at 1.08% of total traffic and ChatGPT at 87.4% of AI referrals. The denominators differ. These sample averages do not establish your site’s share, lead quality or a continuing monthly growth rate.
STEP 1
Record the source evidence
Keep the actual referrer or controlled campaign values, timestamps and classification rule. Leave unsupported source fields unknown.
STEP 2
Preserve the attribution scope
Distinguish the first observed entry from the current session. Follow consent and retention rules; do not promise that a cookie will persist across every visit or device.
STEP 3
Pass through with provenance
On a valid lead submission, carry permitted source fields into CRM with the lead identifier. Distinguish observed, self-reported and inferred values, and deduplicate repeated submissions.
STEP 4
Evaluate qualified outcomes
Compare the same outcome definition, period and eligible cohort. Change scoring only when your own evidence supports it; do not automatically award points for an AI label.
Chapter takeaway
Traceable source fields make a lead easier to investigate. They do not make a source label the lead’s complete history or prove incremental revenue.
CH.05
Use self-reported discovery as a separate signal
A discovery question can record what a person remembers. It can add context when technical data is incomplete, but memory, wording and nonresponse affect the result. Keep the question voluntary and the interpretation limited to respondents.
A person may remember an assistant recommendation but later arrive through Google or a saved link. Both records can describe different moments in the same journey. Agreement is useful context; disagreement does not directly measure hidden AI traffic or prove that one record is wrong.
How to ask
Ask a clear, optional question
Use a short discovery question at an appropriate point in the existing flow. Record the question version and unanswered responses; do not assume any timing guarantees accurate recall.
Options
Offer balanced choices
If you collect structured answers, include relevant channels and an open response. Avoid prompting people to select AI or forcing a single source when several influenced them.
Reconciliation
Compare matched records
Compare technical and self-reported fields for the same responding leads and period. Keep the response denominator and missing data visible; do not subtract two unrelated percentages.
Feedback loop
Retain both accounts
Keep the original response beside the technical record in CRM. Resolve clear collection defects while preserving uncertainty and the possibility of multiple touchpoints.
Chapter takeaway
Self-report adds a person’s account. It does not turn an incomplete attribution system into a complete census.
CH.06
A practical review of your existing measurement setup
Start with the reporting and collection you already use. Review definitions, a small set of known journeys and CRM field handling before adding infrastructure. Setup time depends on access, consent configuration and the existing implementation.
STEP 1
Inspect current channel reporting
Check AI Assistant and the session, user or event scope you need. An optional custom channel group can answer a specific reporting question; it is not required to make the default channel exist.
STEP 2
Validate controlled links
Use campaign parameters for genuine distributions whose source you control. Do not label every internal CTA, documentation link or llms.txt link as AI traffic: ordinary readers can use the same links.
STEP 3
Follow a known test journey
With appropriate consent, inspect the recorded source through the collection and CRM path using a nonproduction fixture. Confirm one successful submission maps to one lead; record what remains missing.
| Check | Evidence of completion | Unresolved limit |
|---|---|---|
| Channel definition | Scope, source and medium are documented | Rules and source lists can change |
| CRM pass-through | A permitted test record retains the observed fields | Missing origin is not reconstructed |
| Qualification | A defined outcome and deduplication rule are used | Source alone does not establish quality |
| Self-report comparison | Matched respondents, missingness and wording are recorded | Recall and selection can differ from the full population |
Tenten’s measurement review checklist; it does not promise a fixed setup time or complete attribution.
Leave the next reader a traceable record
Show what you observed, how you classified it and what you could not determine. That record supports better decisions than assigning every unexplained visit to a promising new channel.
Sources & methodology
Reviewed October 5, 2026. Averi’s April 14 account is a single-site publisher report with inconsistent GA4/Fathom wording; 57% is not a market estimate or a measured AI share. https://www.averi.ai/blog/57-of-our-traffic-is-direct-unknown-what-it-means Conductor’s report (updated July 6, 2026) uses 1,215 enterprise customer domains for U.S. traffic analysis, May–September 2025; 1.08% uses all traffic as denominator, while 87.4% uses AI referrals. Its separate 13,770-domain citation analysis is not the traffic sample. https://www.conductor.com/academy/aeo-geo-benchmarks-report/ Current GA4 channels, Direct and Unassigned are defined by Google. https://support.google.com/analytics/answer/9756891 https://support.google.com/analytics/answer/15258820 Optional custom groups do not replace missing source evidence. https://support.google.com/analytics/answer/13051316 Server-side event handling, page_referrer and consent behavior follow the official documentation. The conclusion that absent evidence cannot be reconstructed from generic request metadata is an implementation limit, not an attribution recovery benchmark. https://developers.google.com/tag-platform/tag-manager/server-side/intro https://developers.google.com/tag-platform/tag-manager/server-side/common-event-data https://developers.google.com/tag-platform/tag-manager/server-side/consent-mode The CRM, self-report and review steps are Tenten recommendations, not measured client outcomes or a guarantee of ten-minute setup.
© 2026 Tenten GEO · Feel free to cite this whitepaper, with attribution.
Discuss your AI search priorities
Request a 30-minute benchmark interpretationSend a booking request. We’ll email you to confirm a time within 24 hours.