Measurement / framework / Free to use

AI Referral Enquiry Evidence Framework

Separate assistant sessions, confirmed submissions and genuine enquiries before connecting analytics to your CRM.

Growthcraft Editorial · 2026-11-02. AI-assisted research and implementation. Examples are synthetic; Akshay's personal review is not claimed.

Copyable template

Select and copy the complete template below. With JavaScript enabled, you can edit, copy and download it in the interactive workspace.

AI REFERRAL ENQUIRY RECONCILIATION — working sheet
Decision and owner:
Reporting period, timezone, extraction time and expected processing delay:
GA4 property/report, session channel, source/medium and landing-page scope:
CRM form or submission log, matching method, permissions and review owner:
Definition of a genuine qualified enquiry:
Known test/internal/duplicate exclusions and how verified:

MEASUREMENTS — KEEP SEPARATE
1. Observed AI Assistant sessions: [count or unavailable]. This is a session measure, not a count of all AI-influenced people.
2. Search Console AI-feature impressions: [count or unavailable]. This is an impression measure, not extra sessions. Do not add it to the overall Search total.
3. Confirmed non-test submissions with a supported AI-source rule: [distinct submission count or unavailable].
4. Those same submission records reliably matched to CRM evidence: [count or unavailable].
5. Matched submission records reviewed as genuine qualified enquiries: [count, zero or unknown].

ONE ROW PER DISTINCT SUBMISSION (USE SYNTHETIC OR APPROVED PSEUDONYMOUS REFERENCES)
Permitted reference | confirmed submission time | source evidence and scope | match status | qualification | reviewer | uncertainty / next check
Use one stable permitted submission reference for the join. A contact's current source property is not a submission log. If no reliable approved join exists, do not guess from names, email or timestamps; report the systems separately.

CALCULATIONS WHEN THE MATCH IS VALID
Unmatched records = confirmed in-scope submissions − matched submission records.
Match coverage = matched / in-scope submissions × 100, undefined if the denominator is zero.
Confirmed-qualified share of matched = reviewed qualified / matched × 100, undefined if the denominator is zero or qualification is unknown. It is preliminary when any matched submissions remain unreviewed.
An unmatched record may be delayed, absent, duplicate or unlinkable; investigate before drawing a conclusion. Do not call it a lost lead.

SYNTHETIC CHECK
40 in-scope submissions; 30 matched; 6 confirmed qualified. The observed match coverage is 75%, with 10 unmatched. Six of 30 matched submissions are confirmed qualified: 20% of matched. Any unreviewed matched records and the unmatched group remain unknown. These are not conversion rates for all assistant sessions.

NEXT ACTION
Evidence gap | named role | next review time | decision held or permitted:
Review actual enquiry content only in the authorised CRM. Record a dated snapshot and later corrections. Respect consent and retention; do not place identifying information in an AI prompt.

The question a dashboard cannot answer alone

A session from an AI assistant is evidence of a measured visit. It does not tell you whether that visitor submitted an enquiry, became a customer or would otherwise have found you. A CRM contact source describes a contact, not every form submission. This framework gives a marketing operations owner a repeatable way to keep those observations apart and investigate a permitted match.

When to use it

Use it for a completed reporting period when you have consent-aware analytics, a reliable confirmed-submission source and permission to review CRM evidence. The aim is to describe what you observed and identify where evidence is incomplete. If there is no shared submission-level key or no approved basis for joining records, stop at separate totals. Do not reconstruct identities from personal details.

Five decisions to make before presenting the numbers

Write the business question, exact reporting dates, timezone, processing delay and ownership. Set the source rule before looking at outcomes.
Keep GA4 session-level AI Assistant traffic, Google's AI-feature impressions and individual form submissions in separate rows. Check whether filters, exclusions and attribution scopes match their descriptions.
Inventory distinct confirmed submissions in the agreed cohort. Remove known internal tests and duplicates using documented evidence. A button click is not a sent message.
Match only on an approved reliable submission reference. A contact's current original or latest traffic source is supporting context, not a record of every submission.
Review genuine qualification with an agreed definition. Record unknowns and an owner for every unresolved discrepancy. Revisit after the expected processing window.

Worked example

In a synthetic week, 40 distinct confirmed submissions meet the source rule. Thirty match CRM evidence and six of those are reviewed as qualified. Coverage is 30 ÷ 40 = 75%; ten records remain unmatched. The qualified share is 6 ÷ 30 = 20% of matched submissions. It is not 20% of visitors, and those ten unmatched records are not automatically lost enquiries. Use the calculator to repeat the arithmetic.

Failure modes and handoff

Different reporting windows, a changed channel definition, consent gaps, missing referrers, repeat submissions and contact merges can make a seemingly tidy comparison unreliable. Preserve the original extract date and the specific question. If one result depends on a missing join, withhold that result and give the next owner a concrete check. The review prompt can draft an anonymised note, but a person must verify it.

Sources and limits

Checked 2 November 2026: Google Analytics release notes describe the AI Assistant channel introduced on 13 May 2026. Google's channel definitions distinguish session, first-user and event scope and place Google's AI Overviews and AI Mode in Organic Search. HubSpot's traffic-source documentation, updated 4 August 2026, describes original and latest sources, AI referrals and tracking limitations. These vendors do not endorse this original reconciliation method. Examples are synthetic.