Findings · 2026-08-29 · By VSNARY | Emmanuel Orta
93% of our arrivals sent no referrer
Across five sites and 1,104 recorded arrivals, 1,026 carried no referrer at all. Every claim about where AI traffic comes from is being made on the remaining seven per cent.
Over thirty days our beacon recorded 1,104 arrivals across five sites we operate. Of those, 1,026 sent no referrer at all — 92.9%. Everything anyone could say about where those visitors came from rests on the remaining 78.
The identified 78 break down as 48 from search, 14 from social, 11 from other sites, and 5 from AI assistants.
Unknown is not direct
The convention in most analytics products is to file a missing referrer under direct, as though the visitor typed the address in. That was a fair guess fifteen years ago. It is not one now. A referrer goes missing when a link opens inside an app webview, when a privacy setting or extension strips it, when a page declares a strict referrer policy, when traffic crosses between differently configured origins, or when the visitor genuinely did type the address. Those are five different events reported as one, and only the last is direct.
We label it unknown, because that is what it is. It is the largest single category in our data, and dressing it as direct would make every other number in the table look more certain than it is.
What this does to AI-traffic attribution
A dashboard telling you AI assistants are some percentage of your traffic is computing that percentage over the visitors it could identify. Here that is seven per cent of arrivals. As a share of identified traffic, AI is 6.4% of what reached these sites. As a share of all arrivals, it is 0.45%. Same five visits, two numbers, a factor of fourteen between them, and the whole difference is a denominator nobody stated.
The bias does not cancel out, and it does not point in a flattering direction. Assistant apps open links in in-app browsers, which is exactly the context where referrers go missing most often. The traffic most likely to be absent from the identified seven per cent is the traffic these tools exist to count.
Which makes our own AI number a floor
Five AI arrivals in thirty days is what we can demonstrate. The real figure is that or higher and we cannot say by how much, because the evidence that would settle it is the evidence the browser did not send. Publishing 6.4% would be arithmetically defensible and substantively misleading. We report the five, we report the 1,026, and we leave the gap visible.
Where a referrer does arrive it is worth having: a ChatGPT visit landed on a tree-removal page, a Claude visit on a fencing homepage, a ChatGPT visit here on a post about a Google redirect. Five specific arrivals with specific landing pages is a real observation. A percentage extrapolated from five is not.
Three rules that cost nothing to adopt
State the denominator on every share you publish. Report an absent referrer as absent instead of assigning it a source. And keep what a visitor did separate from what you would like it to mean: our beacon recorded fifty on-page actions across these sites — phone taps, email taps, form submits — and not one of them is a lead. A tapped number is not a connected call.
What this measurement cannot do
It is a JavaScript beacon, so it sees people and never crawlers: GPTBot does not run scripts, and nothing here describes what any AI crawler fetched. Five sites over thirty days is a small sample, four local-service businesses and one technical SaaS, which is not a random draw from the web. Ad blockers and failed beacon delivery remove a further unknown slice before we count anything. And a referrer names a host, not a conversation — it can say a click came from an assistant, never which answer produced it.
We would rather publish those limits than a share that survives only because nobody asked what it was divided by.
Every figure above came out of this scanner.
Point it at your own domain and see the same measurements, free.