Measurement · 9 min ·
AI Search Traffic Attribution: A Practical Model
Why click-based attribution breaks
Classic web attribution assumes a click. AI search routinely produces influence without a click — a user reads the synthesised answer, internalises the recommendation and either acts later through a different channel or never produces a measurable session at all.
Forcing AI search into a click-based attribution model will always understate its contribution. The fix is not better referrer parsing alone; it is a model that explicitly accounts for impression-led influence.
The three-layer model
- Layer one — citation share of voice across your target prompts. The leading indicator that you are in the answer at all.
- Layer two — on-site signals from AI sources: referrers, user-agents, AI-tagged campaigns and direct sessions following citation spikes.
- Layer three — trailing macro signals: branded search volume, direct traffic baselines and pipeline contribution by self-reported source.
How the layers reinforce each other
When a content cluster ships and your prompt-level citation share rises, layer two should follow within days as a small click lift. Layer three follows within weeks as branded search and direct traffic rise. A change in layer one without a corresponding move in layers two and three suggests the citations are present but not commercially valuable; the inverse suggests measurement gaps rather than missing influence.
This is why all three layers belong on the same dashboard. Looking at any layer in isolation invites the wrong conclusion.
Self-reported attribution
Self-reported attribution — asking buyers 'how did you hear about us?' on the lead form — has become the highest-signal way to capture AI search contribution at the revenue layer. The answer 'ChatGPT' or 'Perplexity' is unambiguous in a way that no referrer ever will be again.
Treat self-reported attribution as the ground truth for revenue attribution and use the upstream layers to explain and optimise it. The combination is more credible than any deterministic model could be on its own.
What to report and how often
- Weekly — citation share of voice by engine and competitor.
- Weekly — AI-source referrer and user-agent traffic.
- Monthly — branded search volume and direct traffic trend.
- Monthly — self-reported attribution share from new pipeline.
- Quarterly — incrementality tests on specific clusters to validate causation.
Frequently asked questions
Is multi-touch attribution still useful?
Yes, for clicks. But AI search adds a heavy impression-led layer that multi-touch alone will miss. Pair it with citation-level measurement and self-reported attribution.
How do I justify AEO investment without clean clicks?
Lead with citation share of voice and self-reported attribution. Both are defensible, both are easy to communicate, and both correlate well with pipeline over time.
Does Hertz handle this?
Yes. Hertz tracks citation share of voice across every major engine and lines it up against your referrer, user-agent and self-reported data so all three attribution layers sit in one place.