Compare referral cohorts with clear outcome definitions
Keep identified AI referrals, other referrals and direct or unknown visits in separate cohorts with explicit outcome definitions. Show counts beside the illustrative rates and follow demo requests to qualification separately. A later direct visit cannot be reassigned to AI merely because that earlier influence seems plausible.
Worked example — illustrative data.
| Cohort | Illustrative visits and outcomes | Interpretation |
|---|---|---|
| Identified AI referrer | 20 visits; 2 orders | 2 orders per 20 visits; illustrative ratio |
| Other identified referrals | 80 visits; 4 orders | Different mix; not a controlled comparison |
| Direct or unknown | 100 visits; 3 orders | Do not redistribute to AI |
| Qualified demo path | 5 requests; 1 qualified | Keep request and qualification separate |
State the cohort dates, referrer rule, device scope and conversion definition. Check whether events and order identifiers reconcile before calculating rates. Keep returns, repeat visits and qualification lag visible in the report. For a SaaS site, trace the request to its later qualified status rather than treating every form submit as pipeline. Where the sample is small, show counts beside rates and avoid a causal story that the data cannot support.
Give qualification its own follow-up window
For the illustrative five demo requests, record how many have been reviewed and how many are still pending. Do not classify an unanswered request as unqualified. Set the review date before comparing referral cohorts, then show orders, refunds and qualification separately. Compare cohorts with similar follow-up time so a recent enquiry is not judged against a fully reviewed older one.
Decision: Use the cohort to inspect landing-page fit and downstream quality. Do not change acquisition spend or declare that AI visitors convert better from this illustrative table.
Later direct visits do not resolve acquisition history: An AI-assisted buyer may later return directly or use a different device. That possibility does not justify assigning unknown orders to AI. Describe the attribution limit instead.
What if only one referral becomes a sale?
Show the count, the cohort size and the observation period. One sale can justify checking that buyer’s landing path, but it cannot establish that the channel converts better. Retain unknown sources and differences in buyer intent.
Continue with Guides for stores, software teams and local businesses, Conversion tracking for orders and qualified leads and Measure revenue from organic store landing pages.
For purchase and refund event collection, see Google Analytics ecommerce event documentation (checked November 2026).