A citation is not a visit, and a visit is not a paying customer. Keep four separate receipts before claiming that AI search produced revenue.
Four receipts, four different questions
| Receipt | Evidence | What it cannot prove |
|---|---|---|
| AI answer | Exact prompt, provider, time and cited URL | Anyone saw or clicked the answer |
| Crawler request | Verified bot identity, path, status and time | A human visited or the page was cited |
| Human referral | Observed landing and referral source after internal/bot filtering | The specific prompt or answer that caused it |
| Payment | Verified provider event, opaque account link and attribution window | Causality or complete profit |
A reproducible reporting example
Suppose a fixed panel records 2 exact target-page citations in 40 answers. Your edge records 90 verified AI crawler requests. Separately, analytics records 6 external human sessions with an AI referrer, and one of those visitors later links to a verified $20 payment. Report four numbers: 2/40 panel citations, 90 crawler requests, 6 referred sessions, one linked payment. Do not report 90 visitors, a 2/40 click-through rate, or $20 of incremental revenue caused by AI.
If the payment window differs from the visit window, say so. If refunds or model costs are incomplete, contribution profit stays unknown. These example numbers are fictional arithmetic, not Grow or Die results.
Build the audit row
observed_at, evidence_type, source, landing_path,
opaque_visitor_id, opaque_account_id, window_start, window_end,
receipt_id, attribution_statusUse separate datasets for answer observations and customer activity. Do not put prompt text, emails, full query strings or customer identifiers into a public case study. Store a stable internal receipt ID privately; publish aggregate methodology and synthetic examples.
A practical seven-day check
- Fix a small unbranded prompt set and keep provider, locale and timestamp consistent.
- Verify the target page is public, indexable and useful without JavaScript.
- Log crawler identities separately from human sessions; a claimed user agent alone is not verification.
- Link the visitor to an account only when your application observes that identity.
- Join signed payment evidence and complete AI cost for the same account and currency.
- Report unknown and unattributed rows explicitly; compare equivalent windows.
Start with AI traffic analytics, then inspect answer visibility. Use the cost per successful task tool when the acquired user runs an agent.
What this evidence can and cannot prove
This checklist separates observable associations from causal claims. Direct or privacy-stripped referrals stay unattributed; neither a search submission nor a crawler hit proves citation or human traffic.