# How to Distinguish AI Referrals from Direct Traffic By AI Visibility editorial team AI-assisted article. A Dubai business sees fewer sessions attributed to AI assistants and more sessions labeled Direct. Both changes deserve a closer look. The business can inspect which acquisition signals GA4 recorded, test whether redirects preserve a source tag and check whether the reporting definition changed. The totals alone cannot tell the team how many Direct sessions started with an AI answer. Keep two questions separate. First, which recorded sessions carry evidence of an assistant referral? Second, where could the acquisition evidence have disappeared? The first gives you a number you can report. The second gives you an investigation to run. Keeping both in the same record helps the team explain an attribution gap without assigning a source to visitors whose journey it cannot reconstruct. Read Direct as a reporting classification Google Analytics describes (direct) / (none) as traffic without a clear referral source. A visitor might type your URL, use a bookmark or open a link in a context that does not provide usable source information. Google also describes missing tags and redirects as possible contributors. The label is a useful count of how sessions were classified in the report. It supplies no distribution of the behaviors that produced that row. Google Analytics guide to Direct traffic (https://support.google.com/analytics/answer/15258820?hl=en) Compare at session scope, using the same dates and filters. A session classified as Direct and an event with unavailable source information are different observations. Keep the Session source / medium and Session default channel group dimensions with the report. If a channel classification moved because a rule changed, inspect the original source and medium before concluding that arrivals changed. Google's Traffic acquisition report documents those session dimensions. GA4 Traffic acquisition report (https://support.google.com/analytics/answer/12923437) Leave other gaps visible under their own labels. An unassigned channel, a missing dimension value, and a Direct classification can have different causes. A visit that your implementation never records is absent from the report altogether. Treat the report as a view of collected activity under a named definition, with limitations in both classification and collection. Look for evidence at more than one point in the visit A click can carry a browser referrer, a parameter in the destination URL, both signals or neither. Browser referrer policy controls what referrer information is sent. An origin can be available without the referring page path. A policy that omits the referrer leaves an independent URL parameter untouched by that policy, although a redirect or other processing can still drop it. Keep these fields separate when you investigate a visit. W3C Referrer Policy (https://www.w3.org/TR/referrer-policy/) OpenAI's API reference shows examples of citation URLs containing utm_source=chatgpt.com. Use that as an example of a source tag to inspect. The examples illustrate a URL format. Your own evidence determines whether a particular arrival carried the tag and whether your implementation recorded it. A saved link with the same parameter can also be shared elsewhere, so a tag supports a source classification under your rules. It cannot establish which question was asked or whether the visitor clicked inside the original answer. OpenAI API reference example (https://developers.openai.com/api/reference/resources/responses/methods/create) Store the evidence at the point where you actually saw it. For a controlled check, record the initial URL, the final URL after redirects, whether a referrer was present at the destination, and the resulting analytics observation if you can isolate it. The first request may carry a tag while the final landing page has lost it. A classification found in GA4 is separate evidence from the browser trace. Treat the match between them as something to verify. The official Home Box UAE Instagram publication below illustrates another entry point a customer might encounter before arriving at a business website. Its existence establishes where the brand published that material. Even when the destination of an untagged visit is a page featured in a public post, the landing page alone cannot establish that the person came from that post. The same limit applies to pages cited by AI answers. View the selected official Instagram post (https://www.instagram.com/homebox_ae/p/DejgCPQNlMe/) Previously verified official Home Box UAE publication. It illustrates a possible discovery channel. Test an owned path before interpreting a chart Choose a landing route you control and an approved test environment or reporting procedure. Use a unique test source value in a tagged URL. Record the timestamp, device or browser context, consent state if applicable, and expected path. Follow that URL through the same kinds of redirects customers encounter, such as a change from a short URL to the full branch page, a language redirect or a hostname migration. Check whether the final URL still carries the tag and whether the intended measurement event receives the expected information. Repeat with a comparable untagged navigation so you can inspect the absence of the tag. If your test starts from a page you control, you can also compare a link that permits a referrer with one that intentionally omits it. Use consistent conditions and record what differs. That exercise tells you which signals survive the path you tested. It does not measure how often a third-party app uses that path. Keep diagnostic visits out of business-performance counts using your team's established test procedures. When a tag disappears, locate the first step where the destination no longer carries it. A URL redirect that replaces the query string requires a different fix from a collection setup that reads the landing page after a client-side change. If the final URL and expected signal survive but the report does not show a corresponding test row, inspect the collection and reporting steps before changing the channel rule. Test data may need time to appear in the report you normally use. Check common journeys before calling the issue resolved. A Dubai business might have different redirects for its English and Arabic pages, or a booking link might send users through a separate host. Establish which path was affected, when the fix was verified, and which periods could contain incomplete attribution. Avoid retroactively assigning earlier Direct visits to an assistant. A repaired path confirms the observed failure and recovery in testing. It leaves the source of earlier unidentifiable sessions unknown. Reconcile a suspected shift without estimating hidden visits This constructed example uses two periods with the same number of recorded sessions and a stable classification rule. The assistant row falls by 16 and Direct rises by 16. These counts are intentionally simple so the attribution problem is visible. Recorded session group Earlier period Later period Change Identifiable AI assistant 40 24 −16 Direct 160 176 +16 Other recorded traffic 300 300 0 Total 500 500 0 The identifiable share changed from eight percent to 4.8 percent, a decline of 3.2 percentage points. Direct changed from 32 percent to 35.2 percent. The equal movements are compatible with several histories. Sixteen previously identifiable assistant sessions could lose the signal while overall visit behavior stays constant. Alternatively, assistant visits could fall by 16 while unrelated Direct visits increase by 16. Both hypothetical histories generate exactly the same table. If a redirect test demonstrates tag loss after a release, the finding strengthens the case for a measurement problem on the tested path. It does not tell you how many real sessions used that path or would otherwise have carried an assistant signal. If a business later gains a separate reliable dataset linking some affected visits to their entry paths, it can document what that dataset verifies and leave the remaining cases unknown. The traffic totals alone supply no such link. The safest report language gives the observed changes and the result of the investigation separately. For the fixture, the analyst could say that identifiable assistant sessions fell from 40 to 24 while Direct rose from 160 to 176. If an owned-path check found a tag lost on a redirect, the report could name the affected route and verification date. It should still leave the number of historical assistant visits in Direct unspecified. Use landing-page patterns to decide where to investigate A sudden change on one landing route can point you to a deployment, redirect or tag change. Compare the timing with release logs and check whether other known traffic sources show a similar shift on that route. If the same redirect loses source tags from several test links, the evidence points to a broader issue on the destination path. If only one source disappears from a chart, inspect the exact source values and the grouping rule used in both periods. Keep hypotheses and verified findings in separate fields. “Traffic shifted to a page often cited by assistants” is a pattern. “The booking redirect discarded the test source parameter in two documented checks” is a verified observation about a specific route. A change in the split between device groups or language pages can help prioritize where to repeat the check, but a segment label still supplies no origin for a Direct session. A customer's answer to “How did you hear about us?” can add useful evidence at the enquiry level. Record it as customer-reported discovery. Unless a reliable link exists between that response and a particular recorded visit, the response belongs in the enquiry log rather than in a revised GA4 source total. The existing AI Visibility enquiry guide sets out a way to record reports from forms, phone and WhatsApp alongside the traffic figures. AI Visibility referral-to-enquiry guide (/article/how-do-you-measure-whether-ai-search-brought-you-anything-from-ai-referral-to-enquiry/) Where the report exposes distinct source values, compare them before aggregating. A single renamed source can appear to vanish if the older rule matches only its former spelling. Conversely, a source label can stay the same while a site changes how a redirect handles tags. The first problem can be visible in the session export. The second may require evidence from a navigation test. Keep a dated snapshot of both the report and the rule so the team can reproduce the comparison even after it updates the dashboard. Close the measurement incident with evidence A reporting incident can close when the team has checked the classification rule, reproduced any collection failure it found, and verified the repair where applicable. Give the historical date range its own status. If the lost signal cannot be reconstructed, the historical gap remains unknown after the fix. The following month can resume the same documented classification while showing the break in comparability. Put the tracked assistant sessions, Direct sessions and total sessions in the monthly note. Add the known path issue and any repair date. A sentence about what the test could observe helps the reader distinguish a channel count from a broader story about where customers discovered the business. AI Visibility monthly reporting guide (/article/what-should-a-monthly-ai-visibility-report-contain/) Referral evidence record Record the property and reporting period, session dimensions, source definition, date of any tagging or grouping change, original and final test URLs, observed referrer or parameter, and the analytics outcome where available. Preserve the date of the repair verification and identify which periods may be affected. Keep the historical number of hidden AI visits blank when there is no evidence to measure it. AI Visibility is published by Lunasol, which offers GEO and AEO services. The reporting table is constructed to demonstrate the attribution problem. About AI Visibility (https://aivisibility.ae/about/)