Measure identifiable arrivals as referrals
AI referral traffic is the set of measured visits for which your analytics system captured an assistant or search product as the traffic source under a defined attribution rule. It is not the number of times your site was cited. A citation can be displayed without a click, while a visit can arrive through a copied link, an app redirect, or privacy-preserving flow that removes the original referrer.
Report these as separate datasets: referral sessions from analytics, visible citations from a reproducible observation sample, and crawler requests from server logs. Joining them into one “AI visibility” number creates precision the evidence does not support.
Define the referral segment before reading the trend
List the exact referring hostnames you intend to classify and the date each mapping became active. Use the full hostname rather than a loose string such as ai, which will create false matches. Preserve raw source and medium values in a controlled report, then derive a named “observed AI referrals” segment so a mapping change does not silently rewrite history.
Check the collection path:
- Verify the analytics tag fires on public landing pages under the site’s consent rules. Google’s GA4 collection documentation describes the event-based collection model; your implementation and consent state determine what enters the property.
- Exclude office, test, preview, and monitoring traffic using the site’s normal internal-traffic process.
- Review redirects between the referring product and landing page. A chain through an app, link service, or HTTP-to-HTTPS hop can change or remove the referrer.
- Record the property timezone, session definition, attribution settings, and reporting identity. These choices affect how totals are grouped and compared.
- Report sessions, engaged sessions, landing pages, and clearly defined outcomes such as signup completion. Use the same date and segment rules for comparison periods.
Google’s current Analytics Data API schema exposes session-scoped dimensions such as sessionSource, sessionMedium, and session source/medium fields in its official dimensions reference. Use session-scoped fields for a session report; do not mix them casually with first-user or event-scoped attribution.
State what the segment misses
Direct and unassigned traffic may include AI-originated visits, but it also includes bookmarks, typed URLs, privacy loss, untagged messages, and many other sources. Do not reallocate a rise in direct traffic to AI without first-party evidence. Likewise, a known referral hostname identifies the immediate source reported to analytics, not the prompt, answer, or exact citation that preceded the click.
Keep manual citation observations in their own table with product surface, prompt, locale, date, visible URL, and result. The AI citation tracking methodology explains the sampling limits. Compare citation and referral trends descriptively only when the windows and page sets align.
Audit the segment after analytics configuration changes. A new consent default, cross-domain linker, referral exclusion, or session timeout can alter totals while visitors behave exactly as before. Keep a dated mapping table for recognized hosts and test one known referral in a staging or controlled environment where policy permits. Do not use a screenshot of a dashboard as the only record; export the dimensions, filters, timezone, and report date range that produced the number.
If an assistant strips referrers, a campaign URL you control can provide a different observation, but it must be labeled as tagged traffic rather than organic assistant referral. UTM parameters identify the link creator’s classification, not whether the assistant selected the page independently. Preserve that distinction in dashboards and editorial reports.
Explicitly hypothetical monthly report
Imagine a fictional SaaS company defines three exact assistant referral hosts on July 1. In a hypothetical July report, GA4 shows 140 sessions in that segment, 82 engaged sessions, and 11 completed trials. Twenty additional sessions contain campaign tags the company placed in links it controls; those are reported separately from passive referrers because the collection mechanism differs.
The team also runs a fixed 30-prompt citation sample and observes its domain in four answers. It does not divide 140 visits by four citations or call that a click-through rate: the citation sample is incomplete, many prompts were never tested, and the referral sessions cannot be tied to those four displays. A rise of 90 direct sessions remains unassigned.
The report’s conclusion is limited but useful: 140 measured sessions arrived with the defined referral signal, and 11 completed the stated trial event under the property’s settings. The mapping, date window, and consent limitations travel with the number.
FAQ
Can I count direct traffic as AI traffic?
No. Direct traffic lacks enough source evidence to make that assignment.
Does a visible citation prove someone visited the page?
No. A visit requires a measured arrival; citation display and referral are separate observations.