Adobe Experience Cloud
Adobe Analytics cannot name every AI referral: measure the unknown
Design honest reporting for visits from AI assistants when browsers and apps omit referrers or campaign tags.
October 28, 2026
A marketer sees visits from Copilot or Claude, yet some arrive with an empty document.referrer and appear as Direct. An Adobe Analytics community discussion reports that other clicks from the same services do carry a referrer or a UTM parameter. That mixed result matters: an Analytics rule cannot recover a source the navigation never transmitted.
Split observed AI traffic from unidentifiable traffic
Build a small test matrix by assistant, web or app surface, browser, and link-opening method. For each click, record the landing URL, request referrer, JavaScript referrer, campaign parameters, and the resulting Adobe Analytics hit. Use fresh sessions and avoid accidentally reusing prior campaign attribution. Label visits with a known conversational-AI referrer or a deliberate campaign parameter as observed AI traffic. Leave the rest unknown; do not relabel all Direct traffic as AI.
Inspect the channel rules before reporting
Adobe's Marketing Channel processing-rule documentation includes a “Referrer Is Conversational AI” condition. It can classify hits where that referrer is actually present. Review rule order and tracking-code handling so a broad rule does not swallow a more specific tagged visit. Adobe also notes that channel-rule changes are not retroactive; record the effective date in reporting.
Give the business a bounded answer
Report the verified count as a lower bound, with a separate unknown-referrer bucket and the test matrix that explains the gap. If your own links are placed in an AI experience you control, tag them consistently. For third-party answers, you cannot guarantee a referrer or UTM. Compare landing pages and conversions only within identifiable cohorts; otherwise the channel story becomes more precise than the data.
Adovo’s view
Attribution is an observation system, not a truth machine. When a source is not sent, the honest result is a measured blind spot. A useful dashboard shows what is known, how it was classified, and where the evidence stops.