AI traffic attribution without false precision

Use an attribution evidence ladder

The evidence determines the strength of the language before a model assigns credit.

An observed AI referral followed by a same-session outcome supports the clearest simple claim. First-party account continuity can associate a later outcome when the identity and lookback rules are disclosed. Self-reported discovery and modeled contribution can add useful context but do not become observed referrals.

Unknown sessions remain at the bottom of the ladder. A visibility increase or matching date range can motivate investigation, but it cannot assign provider-level revenue to visits that contain no source evidence.

Write the attribution rulebook first

A reproducible policy prevents the winning channel from rewriting the rules after the result appears.

Define session boundaries, identity keys, consent behavior, source precedence, cross-domain handling, lookback windows, exclusions, currency and value rules, and the events that qualify as outcomes. Specify how returns through paid search, organic search, email, or Direct affect credit.

Version the rulebook and apply it consistently across providers. Preserve touchpoints where lawful and necessary so analysts can compare first-touch, last non-direct, and multi-touch conventions without changing the underlying evidence.

• Outcome dictionary
• Session and lookback window
• Identity and consent rules
• Source precedence
• Cross-domain policy
• Value and refund handling

Choose a model that answers a named question

First-touch, last-touch, and multi-touch models are conventions, not natural facts.

First-touch can describe measurable discovery, while last non-direct can describe the known source preceding an outcome. Position-based or data-driven models distribute credit and require enough clean history to be stable. None reconstructs a referrer that was never observed.

Publish the question beside the result: acquisition, conversion efficiency, pipeline association, or modeled contribution. Do not compare numbers from models with different windows and present the difference as channel performance.

Handle long B2B journeys explicitly

Pipeline and revenue often arrive long after the first measurable AI visit.

Connect an observed visit to a later account only when a lawful first-party identifier supports continuity. Distinguish created pipeline, qualified pipeline, closed revenue, and retained revenue. State whether the report uses a person, account, opportunity, or transaction denominator.

Show the observed entry page and intermediate meaningful actions where available. An AI referral to documentation that later contributes to an enterprise opportunity may be strategically important even if another channel receives final-touch credit.

• Person and account boundary
• Opportunity creation date
• Qualification stage
• Lookback window
• Other known touchpoints
• Closed and retained value

Keep small provider samples honest

A large percentage change can be one additional conversion.

Show visits, conversions, and value beside rates. Use comparable periods and annotate product launches or tracking changes. Avoid ranking providers when the sample is too small to support a stable conclusion.

Where decisions justify it, add confidence intervals or aggregate across a longer window. Continue inspecting landing-page and intent differences because a provider may send fewer but more valuable visitors.

AI conversion rate = observed AI-referred visits with outcome ÷ observed AI-referred visits × 100
AI value per visit = attributed value under the stated rule ÷ observed AI-referred visits

Publish an executive answer with an audit trail

The top line can be concise as long as the evidence classes remain available.

Lead with observed AI-referred conversions and value. Follow with associated outcomes, self-reported influence, and modeled contribution in separate rows. Show unattributed volume and the current coverage note.

Link the result to the source taxonomy, attribution rule version, event dictionary, and page-level path. A reader should be able to understand the decision immediately and inspect the assumptions without requesting a new analysis.

Audit the common attribution failure modes

Most overstatement comes from a small set of repeatable mistakes.

Do not backfill Direct because visibility increased, combine self-report with observed referrals, use different conversion definitions by provider, silently extend lookback windows, count machine activity as sessions, or compare event-attributed revenue with session-acquisition revenue.

Review the rulebook quarterly and whenever identity, consent, checkout, CRM, domain, or channel classification changes. Preserve the date of the change so historical differences remain interpretable.

Evidence note

If two analysts applying the same documented rules cannot reproduce the classification, the attribution model is not operationally complete.