Arrow 
An AI attribution stack should preserve referral evidence, measure meaningful website actions and connect those actions to CRM outcomes through a documented join. Store observed source, self-reported discovery and model-based attribution separately. No source label can recover a visitor's private AI conversation or prove incremental revenue by itself.
What data belongs in an AI attribution stack?
Build around evidence records before choosing software. An answer observation describes what appeared in a particular test. An acquisition record describes a recorded visit. An inquiry record describes a submitted request. An opportunity record describes the commercial process. Each needs an identifier, timestamp and evidence type.
A spreadsheet may be enough to reconcile a small cohort. A warehouse becomes useful when volume and joins require it. Neither architecture creates evidence that was never collected. Begin with the citation tracking foundation, then decide which joins can be supported in your existing analytics and CRM.
Download the blank attribution ledger to document the records you can support. It is a reusable worksheet, not an automatic analytics or CRM integration. Populate it with appropriate internal references and aggregated evidence, without personal information.
| Record | Minimum useful fields | What it establishes |
|---|---|---|
| Answer observation | Observation ID, prompt, surface, time, response, cited URLs | A result in a defined test |
| Acquisition | Recorded source, landing path, time, permitted session reference | An observed website entry |
| Inquiry | Internal inquiry ID, created time, evidence category, qualification | A distinct request and its source evidence |
| Opportunity | CRM ID, linked inquiry IDs, stage dates, amount convention | Commercial progression under the CRM definitions |
How should you collect and normalize referral evidence?
Preserve the available source values before applying your reporting categories. OpenAI says its ChatGPT search referral URLs include utm_source=chatgpt.com. Inspect a real, permitted test click through your own redirects and forms. Confirm the actual values recorded instead of assuming every arrival will use a particular medium. OpenAI publisher documentation explains this signal.
Maintain a versioned mapping from known source domains or campaign values to an AI-referral category. Keep the original value for auditability. Do not replace an unknown source with AI because traffic arrived after a visibility campaign. Do not place personal information or a private prompt in UTM parameters. For records needed to join analytics and CRM, use an appropriate internal identifier under your consent and data-handling rules.
Which GA4 view answers which attribution question?
Use session-scoped acquisition dimensions to review recorded visits. First-user dimensions answer a different question about initial acquisition, while event-scoped attribution can allocate credit for key events. Google documents these distinctions in traffic-source scopes. State the scope in the report title so readers do not mistake one result for another.
Choose a meaningful, successfully completed action to measure. A click on a submit button can occur without a valid inquiry reaching the CRM. Google describes key events as events important to the business; defining one still does not make it a sales-qualified lead. Reconcile the event against the actual inquiry before using it as a commercial outcome.
- Record the property, reporting timezone, date range and selected source dimension.
- Document the key-event definition and whether repeat submissions are counted.
- Keep acquisition totals separate from credit assigned by an attribution model.
- Record instrumentation changes before comparing one month with another.
How do you join an inquiry to a CRM outcome?
Use a documented, permitted identifier created in your own workflow. For example, a successful inquiry can receive an internal request ID that is also stored on the CRM record. Keep the source evidence attached to that request rather than overwriting the contact's entire acquisition history every time a form is submitted.
Track the method and confidence of every join. An exact request-ID match is different from a salesperson's recollection. Multiple contacts may belong to one opportunity, and one contact may submit multiple requests. Choose whether the commercial report counts requests, accounts or opportunities, and deduplicate at that level. Keep identity details in authorized systems; public dashboards should contain only the aggregates required for the decision.
| Suggested field | Example value | Reason to retain it |
|---|---|---|
| source_evidence | observed_referral | Separates a recorded source from self-report |
| inquiry_id | REQ-EXAMPLE-104 | Links the successful request to its CRM record |
| join_method | request_id_exact | Shows how the connection was made |
| qualified_at | Recorded acceptance timestamp | Preserves when the lead met the criteria |
| attribution_rule | observed_entry_v1 | Makes the reporting convention reproducible |
What does a reconciled attribution example look like?
Fictional example: analytics records 12 successful inquiry events for a source cohort. The CRM contains 10 distinct inquiry IDs after two repeat submissions are deduplicated. Eight match the recorded source evidence, one has only a self-reported AI discovery answer, and one has unknown acquisition. Report eight observed-source inquiries, one self-reported inquiry and one unknown. Do not report 12 unique AI leads.
If three of the eight observed-source inquiries become opportunities, report that progression with the cohort dates. Any opportunity value is pipeline under the CRM's amount convention, not collected revenue. A later direct visit or another marketing touch may also matter. The cohort shows an association supported by recorded evidence; incrementality needs a separate evaluation design.
Which checks should run before a monthly report?
Test a complete path through landing page, consent state, successful form submission and CRM creation. Check redirects, cross-domain steps and embedded forms where they exist. Compare distinct inquiry IDs with counted events, investigate missing joins, and keep test submissions out of outcomes. Recheck after site or form changes.
Publish a coverage note: which sources are identifiable, how many inquiries could be joined, and what remains unknown. Use the lead-journey guide to interpret conversion stages and the CFO reporting memo to turn the evidence into a decision. The measurement hub connects these workflows.
Sources and editorial scope
This is Arrow AI's implementation guidance. Examples are illustrative unless identified as dated observations. Source access and good content do not guarantee a recommendation.
- OpenAI: Publishers and Developers FAQ
- Google Analytics: Scopes of traffic-source dimensions
- Google Analytics: About key events
- Google Analytics: Campaigns and traffic sources
Continue through the GEO evidence library, inspect Arrow GEO's measurement limits, or start an audit.