The AI Attribution Gap: Known Referrals and Missing Sources

An AI-referred visit can reach your site without a usable referrer. GA4 may then record it as Direct, leaving you unable to identify its origin from the session alone.

Direct includes typed URLs and bookmarks, but it also includes visits whose source information was lost. Copying a link out of an assistant answer or opening it in some app browsers can remove that information. A rise in Direct does not tell you how much came from AI, stronger brand recognition, or another source.

This series separates known AI referrals from uncertain influence, then follows the known signals into commercial reporting. Conversion and revenue results vary by site and engine; your own funnel has to establish their value.

What the 70.6 percent estimate can establish

Loamly's 2026 vendor benchmark, covering 446,405 visits, estimates that 70.6 percent of AI-driven traffic arrives without a referrer. It does not publish the classifier, validation set, or false-positive rate needed to audit that estimate. It is a vendor estimate, not an established share of your Direct traffic. The dark AI traffic analysis checks those limits.

Referrer loss is a known mechanism. Its frequency depends on the assistant, browser, link, and user action. Once the browser supplies no source signal, a later analytics report cannot recover that origin merely by relabelling the visit.

What Google's AI Assistant channel measures

Google announced the GA4 AI Assistant channel on May 13, 2026. It provides a useful category for recognized assistant referrals, with a narrower job than measuring all AI-influenced demand.

Google's channel definition uses recognized referrers and explicitly places AI Overviews and AI Mode in Organic Search. Its assistant examples are non-exhaustive; their omission of a platform does not prove that platform is excluded. Check source-level traffic and annotate the reporting change before comparing periods. The GA4 channel analysis covers the practical checks.

Why studies report different shares

Published AI-referral studies put ChatGPT's share anywhere from roughly 62 to 92 percent. Different audiences, collection methods, periods, and definitions can produce different shares. Without equivalent cohorts, the spread cannot prove an instrument is broken or quantify missing traffic.

Use those reports to frame questions for your own data. Keep the study window and denominator beside each figure rather than blending them into one market share.

What can be measured

Recognized referrers provide observed source information. Classifying no-referrer sessions by page depth, timing or geography introduces uncertainty: other channels can produce the same behavior. Earlier dark-social and organic-search analyses show why the method and denominator need checking. The dark-traffic article explains why an estimate needs a validation method and a false-positive rate.

Company identification answers a separate question: which account might be visiting. It does not recover the visit's source. A beacon can supply a denominator for evaluating identification coverage, and reviewed matches can support a precision score. Neither measurement establishes that an unidentified referrer came from an AI assistant. The defense-sector demand case describes the commercial program and its documented outcomes, rather than validating a universal dark-traffic recovery rate.

Following the known signal into revenue

Some studies report stronger conversion among observed AI referrals than among organic visitors. That result does not establish the value of the unobserved visits or prove the assistant caused a sale. Keep observed referrals, estimated influence, pipeline, and financial outcomes distinct. Proving AI Pipeline to Your Board covers the contact-to-deal handoff; the AI commercialization guide covers the broader operating work.

The series

The four published articles:

Why GA4's AI Channel Still Undercounts: recognized referrals, Google AI surfaces, and reporting-period comparisons.
Dark AI Traffic: What the 70 Percent Estimate Can Prove: the vendor estimate, source criticism, and the limits of recovery.
AI Referral Value by Engine: revenue per visitor, conversion events, and conflicting benchmark arithmetic.
Proving AI Pipeline to Your Board: preserving source evidence and reporting attributed commercial outcomes.

Frequently asked questions

Why does AI traffic show up as direct in GA4?

A copied link or some app-browser journeys can arrive without a usable referrer. GA4 may classify that visit as Direct. Loamly estimates a 70.6 percent hidden share in its vendor benchmark, but does not publish enough classifier-validation evidence to establish that share or apply it to your site.

Does GA4 track AI assistant traffic?

GA4 has an AI Assistant channel for recognized assistant referrers. It cannot identify the source of every no-referrer visit, and Google explicitly includes AI Overviews and AI Mode in Organic Search. Platform examples in its documentation are not an exhaustive matcher list.

How do you track AI referral traffic accurately?

Report observed assistant referrers by source and preserve that evidence through the CRM. Keep any inferred no-referrer share separate, with its method and validation limits. Server or CDN logs can show crawler activity, but a crawl does not prove a human referral or recover a missing source.

Does AI traffic convert better than search?

Some site-specific studies report higher conversion for observed AI referrals. The result depends on the site, period, source definition, and conversion event. It does not establish a universal ranking or the conversion rate of visits whose origin is unknown.

What is AI attribution?

AI attribution assigns observed visits and commercial outcomes to assistant sources under a stated method. Estimated influence is a separate claim. Attribution alone does not establish incremental sales.

Which AI assistant sends the most traffic?

ChatGPT leads many published observed-referral studies, but the reported share varies by cohort, period, and method. A range across unrelated studies does not establish a universal market share or prove measurement failure.

Sources

Primary channel definition: Google Analytics default channel group. Vendor estimates: Loamly State of AI Traffic 2026 and SE Ranking's AI-traffic study. Each uses its own cohort and collection method. The linked companion articles retain the source-specific limits.

About the author

Jeff Brokaw is a CMO and Certified Chief AI Officer who builds AI workflows, measurement systems, and the commercial operations around them.