What AI referral attribution can measure
A practical attribution layer starts with session-source evidence exported from Google Analytics 4. Each row should retain the landing page, session source or source / medium, sessions, engaged sessions, key events or conversions, and revenue. Full referrer and campaign fields can provide additional explicit identifiers. IndexHalo classifies a row only when one of those supplied fields names a disclosed AI answer product such as ChatGPT, Perplexity, Gemini, Copilot or Claude.
The result is a measured subset of analytics traffic: sessions that the source data explicitly identifies. It is not an estimate of every visit influenced by an answer engine. Organic search, direct, unassigned, unknown and “not set” sessions remain in their original categories.
Unknown does not mean AI. A defensible report prefers a smaller measured number over a larger number created by redistributing unattributed traffic.
Build the right GA4 export
Use a report with one row per landing-page and source combination. Include the date window in the export or record it at import. Useful dimensions are landing page plus query string, session source, session medium, session source / medium, full referrer, default channel group and campaign. Useful metrics are users, sessions, engaged sessions, key events and total or purchase revenue.
Keep source strings intact. Renaming every answer product to “referral” removes the provider evidence needed for a provider-level view. Conversely, a broad channel label such as referral should never be sufficient on its own to call the row AI.
Reconcile the totals before interpreting them
The provider table and landing-page table must both reconcile to the same identifiable-AI totals. Sum sessions, conversions and revenue across providers, then repeat the sum across landing pages. Compare the identifiable session total with all GA4 sessions to calculate a transparent share. This check catches duplicate exports, mixed scopes and accidental double-counting.
Period movement requires equivalent windows. Compare two dated periods only when their lengths match within a narrow tolerance. If dates are missing or window lengths differ, show the current measurement and withhold the change. Seasonality and campaign changes still require analyst review even when the lengths match.
Keep referral and citation evidence separate
A provider row may have both measured referral sessions and timestamped citation observations. That creates an observable funnel for investigation, not proof that the observed answer caused every session. A provider may cite the site without creating an identifiable visit, or send referral traffic in a period where the fixed observation set did not cite the target.
Those combinations are useful diagnoses. Visibility without measurable referral can trigger a tracking and cited-URL review. Referral without a current citation can trigger a question-set coverage review. Traffic from a provider not yet observed can justify adding that provider to the fixed test set. None of these states should overwrite the underlying analytics or citation denominators.
Turn landing pages into commercial evidence records
Join each AI-referral landing page to the public crawl. Show its GEO score, highest-severity finding, answer-ready questions, material claims, adjacent support and conversion economics. A page with strong revenue and unsupported claims deserves provenance repair. A page with high sessions and zero conversions needs a conversion-path investigation that preserves the answer earning the traffic.
PDFs deserve special treatment. If a PDF attracts AI referrals but converts poorly, publish a canonical HTML evidence hub containing the same key table, scope, method and sources, keep the PDF as a download, and test the commercial next step on the HTML page. Do not remove a useful source asset merely because its conversion experience is weak.
Write actions that a team can implement
Every recommendation should identify the owner, exact landing page, measured baseline, evidence problem, implementation, and repeat test. “Improve this page” is not enough. A useful action reads: preserve the current payback answer, attach the 84-site methodology beside the headline figure, add a qualified consultation step after the assumptions table, and compare conversion rate and revenue per session in the next equivalent GA4 source window.
Rank the queue using actual revenue or sessions alongside evidence severity. High-value pages with crawl failures, factual conflicts, stale evidence or low conversion rates should rise above generic new-content ideas. Protect working pages: a strong revenue-bearing answer should not be rewritten casually to chase a modelled score.
Do not confuse attribution with causality
Session-source attribution reports where an identifiable visit was recorded. It does not establish why the provider returned the link, whether a particular GEO release caused it, or whether the user saw a citation. Controlled before-and-after experiments can strengthen the decision evidence, but they still need stable pages, fixed questions, equivalent measurement windows and disclosed confidence limits.
What the professional export should contain
Export record type, provider, landing URL, state, users, sessions, engaged sessions, engagement rate, conversions, conversion rate, revenue, revenue per session, period changes, separate citation checks and citations, page score, answer-ready questions, material and unsupported claims, action, verification and evidence class. Retain the source import date and period definition so another analyst can reproduce the conclusion.