When a quarter of your customers say an AI assistant sent them and your analytics credits AI with low single digits, neither figure is wrong. They measure different things. The survey captures influence, which is what AI actually does: it shapes the decision, then hands the customer off. Analytics captures the last click, which by then often looks like direct traffic. The gap is not an error to fix. It is a feature of how AI referrals travel, and you reconcile it rather than trust either side alone.
The cleanest illustration comes from Doola, the business-formation company, which found over 25% of customers self-reported arriving via an LLM while tracked attribution showed single digits. A 3-to-5-times gap between what customers say and what the dashboard records. Doola is not sloppy at analytics. The traffic genuinely arrives looking like something else.
Why does AI traffic hide as direct?
Because the referrer often does not survive the trip. When someone reads an answer in the ChatGPT or Gemini mobile app and taps through, the app frequently strips the referrer header. HTTPS redirect chains drop it in transit. Privacy browsers remove it by default. GA4 sees a visit with no source and files it under “direct” or “(not set),” which is analytics-speak for “someone showed up and I have no idea how.”
The scale is measurable. An analysis of 181.6 million GA4 sessions found roughly 22% of ChatGPT sessions and about 32% of Perplexity sessions dumped into the “(not set)” medium, while Claude and Gemini were attributed cleanly to referral. So even the “AI traffic” your dashboard does label is missing a fifth to a third of itself. Other practitioners put the direct-misclassification rate higher still, with work on the GA4 direct-traffic problem describing the majority of LLM traffic landing as direct.
Then there is the traffic that leaves no click at all. A customer asks an assistant about your category, gets a recommendation, and types your brand name into the address bar the next day. No referrer, no AI-labeled session, nothing. That path shows up as direct or branded search, and it is pure AI influence with zero digital fingerprint. Surveys are the only instrument that catches it. In one set of client surveys across 20,000-plus responses, 35% of visitors reported using an AI tool before their visit.
Which number should the board see?
Both, side by side, with the method attached. The survey number is your ceiling: total AI influence, including the invisible path. The analytics number is your floor: AI traffic clean enough to trace. The truth sits between them, and pretending either endpoint is the whole answer is how you either panic the board or lull it.
There is a tell that helps you trust the floor. When an AI referral does pass a clean URL, it sometimes carries a watermark. In our own testing of engine outputs, citation URLs returned by OpenAI’s search surface came stamped with utm_source=openai. That is the provider labeling its own traffic. Where you can catch those parameters in your analytics, you can positively attribute a slice of what would otherwise vanish into direct, which turns a guess into a confirmed count for at least part of the flow.
How do you reconcile the two numbers?
Build the floor first. In GA4, create a channel or segment that catches known AI referrers and campaign parameters, and stop letting them fall into direct. That recovers the clean portion. Then build the ceiling: add “how did you hear about us” at signup or checkout with an explicit AI-assistant option, and treat the delta between survey and analytics as your dark-referral estimate, not as a rounding error to bury.
Watch the segment you just cleaned up, too. If your labeled “direct” traffic converts far above the rate direct traffic should, that lift is almost certainly misfiled AI referrals dragging the average up. Direct traffic converting like a warm inbound lead is a symptom, and the diagnosis is usually AI.
None of this is exotic, and all of it is undone the moment you report a single attribution figure as if it were the truth. Attribution reconciliation is one layer of honest AI visibility measurement, sitting between the sampling that tells you whether AI mentions you and the survey that tells you whether customers noticed.
Open GA4, find how much of your “direct” and “(not set)” traffic spikes alongside AI mentions of your brand, and add the one survey question that catches the customers your logs never will. The reconciled range is the number worth defending. Our free scan checks whether AI assistants can find and cite you in the first place, which is the upstream half of the same question this gap is asking.