Why Were 19 of 22 AI Traffic Sessions Classified as Direct?
Published
This article was first published in Korean by Intrix Lab on September 14, 2026.
Of 22 sessions carrying AI Assistant A's UTM identifier, 19, or 86%, were classified as direct traffic. This was observed in Intrix's own tracking over the most recent 30 days as of 2026-09-12. Journey classification based only on the referrer and inquiry attribution that checked UTM first recorded the same visits differently.
What were these 22 visits?
The measurement period was 2026-08-13 through 2026-09-12. After excluding internal visits and suspicious traffic, 22 sessions carried AI Assistant A's UTM identifier. Of these, 19 had no referrer and were classified as Direct, while 3 were classified as AI because the assistant was identified in the referrer.
Why was the same visit assigned to different channels?
The cause was the order of classification, rather than the data itself. In the classification that determined channels from referrers, the 19 sessions with empty referrers received no AI traffic signal and became Direct. In contrast, the classification that checked UTM first retained those same visits as AI. The same visitor could therefore appear as either Direct or AI, depending on which report was viewed.
Was this only a mobile app issue?
Of the 19 sessions classified as Direct, 11 were on desktop and 8 on mobile. An explanation that mobile apps removed referrers cannot account for the 11 desktop sessions. We did not establish whether desktop apps or referrer policies were responsible.
What we confirmed was limited to the presence of a UTM identifier and the absence of a referrer. We did not interpret the 86% as the effect of a particular device or app policy.
What are the limits of these findings?
The population consists of 22 UTM sessions from AI Assistant A arriving at a single site, Intrix, over 30 days. The sample is small and limited to a specific period, so there is no guarantee that the same 86% would appear on other sites or during other periods. We also did not examine other AI assistants' UTM practices.
Visits carrying multiple identifiers may have been excluded from this count. The two classifications also did not use entirely identical filters, so the 19 Direct sessions and 3 AI sessions cannot be considered a comparison under identical conditions.
What can sites in similar circumstances take from this?
AI traffic counts based only on referrers can have measurement blind spots. This is the extent of what the data supports. When reporting AI traffic, count referrer-based classifications and UTM detections separately and record the difference.
Intrix's free SEO and GEO audit checks a site's search and crawling conditions. AI traffic attribution requires separate verification that considers both referrers and UTM parameters. Intrix does not guarantee specific rankings, citations, or visibility.
Frequently asked questions (FAQ)
Q. Can we reclassify all 19 Direct visits as AI traffic?
These 19 sessions carried AI Assistant A's UTM identifier, so they warrant a separate count. However, keep the original count alongside a sensitivity analysis that includes UTM, without changing the definitions of the existing preregistered measurement midway through.
Q. Is 86% the typical undercount rate for AI traffic analytics tools?
No. This proportion was observed in a 30-day sample from a single site, Intrix, and a single assistant, Assistant A. It cannot be applied directly to other measurement setups.
Q. Did you establish why the referrers were missing?
No. With 11 desktop sessions and 8 mobile sessions, we did not attribute the result solely to a mobile app. The current data establishes only that UTM identifiers were present and referrers were absent.