Why your AI traffic still shows as direct

AI referral traffic gets logged as direct whenever the click reaches your site without a Referer header, because analytics tools have no other signal to name the source with and fall back to the unattributed bucket.

That single mechanic is why teams keep reporting that ChatGPT sends them almost nothing while their sales team keeps hearing "I found you through ChatGPT." The traffic arrived. The label did not.

#Google closed part of the gap

Google Analytics 4 now ships an AI Assistant channel in its default channel group, described in Google's own documentation as the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok. It matches on either of two conditions: the medium is exactly ai-assistant, or the referrer matches Google's maintained list of AI assistants and GA4 sets the medium itself.

This is a real improvement. Until it landed, every AI click that did carry a referrer was scattered across Referral or lumped into Organic Search depending on how you had configured things, and comparing month to month meant maintaining your own regex list.

But read the matching conditions again. Both of them need something to match on: a tagged medium, or a referrer. If neither survives the trip, GA4 assigns the session to Direct, which the same documentation defines as source exactly (direct) with medium (not set) or (none). There is no third rule that recovers an untagged, referrer-less click. It goes in the same bucket as bookmarks and typed URLs.

#Where the referrer actually disappears

Four mechanisms account for most of the loss, and only one of them is anybody's fault.

The link carries rel="noreferrer". MDN defines this attribute as instructing the browser to omit the Referer header and otherwise leak no referrer information. It is a deliberate privacy choice by the linking site, and when it is present there is nothing on your end that can recover the source.

The page sets a restrictive referrer policy. Per MDN, the no-referrer policy omits the Referer header entirely for every request. Worth noting what is not a problem: the modern browser default is strict-origin-when-cross-origin, which still sends the origin on cross-origin HTTPS requests. That is enough for GA4 to name the source. The default is fine. Explicit tightening is what costs you attribution.

The click happens inside a mobile app's built-in browser. Assistant apps often open links in an embedded webview rather than handing off to the system browser. What the webview passes along varies by platform and is not something you control.

The user copies the URL and pastes it. This one is not a bug at all. There is no linking document, so there is no referrer to send. It is genuinely direct traffic by every technical definition, and it is common behavior when someone is reading an answer on one device and acting on another.

#The tagging that survives

OpenAI has been closing this from its side. ChatGPT appends a UTM parameter to outbound links, and in June 2025 it extended that tagging to links in the "More" section of a response, which had previously been standard links without any tracking parameters.

The important property of a UTM parameter is that it lives in the URL string, not in an HTTP header. It survives rel="noreferrer", it survives a restrictive referrer policy, and it survives copy-paste, because the user copies the whole tagged URL. Header-based attribution is fragile in exactly the places query-string attribution is not.

#Is this worth the effort?

By volume, no. By value, probably yes.

Ahrefs studied 3,000 websites and found that 0.17% of the average website's visitors come from AI chatbots, with 63% of sites receiving at least some. The same study found the distribution heavily concentrated: ChatGPT sends 50% of AI traffic, Perplexity 30.7%, and Gemini 17.6%. Those are small fractions, and any measurement plan that ignores that fact will over-invest.

The value side reads differently. Reporting on their own analytics, Ahrefs found that AI search drove 0.5% of traffic but 12.1% of signups over a 30-day window, a conversion rate they put at 23 times higher than traditional organic search. One company's numbers are not an industry benchmark, and the same post carries a caveat worth repeating: their modeling suggests users of AI search click links roughly 75% less often than they do in traditional organic search. The people who do click are a filtered, late-stage group, which is most of the explanation for the conversion gap.

The practical read is that this channel is currently small, high-intent, and badly measured. Mismeasuring a small channel is cheap. Mismeasuring the channel that converts best is how budget gets cut from the thing that was working.

#What to do this week

Confirm the AI Assistant channel is populating. If your property predates it, check that you have not pinned a custom channel group that shadows the default one, and that older custom rules matching chatgpt.com as a Referral are not competing with it.

Stop stripping the parameters. Some CDN and cache configurations drop unknown query strings to improve hit rates, and some frameworks rewrite URLs on redirect. Either will remove a UTM parameter before your analytics tag reads it. Test by loading one of your own URLs with the parameter attached and confirming it appears in a realtime report.

Segment landing pages, not just sessions. The useful output is which specific pages AI assistants send people to, because that tells you what got cited. A sessions total tells you nothing actionable.

Add a self-reported attribution field. A "how did you hear about us" question on signup or contact forms is the only method that captures the copy-paste path, and it is the one segment analytics is structurally unable to see. Treat the gap between that field and your GA4 numbers as your measurement error bar.

Do not expect the buckets to reconcile perfectly. Some portion of AI-driven traffic is unattributable by construction, and a plan that assumes otherwise will burn time chasing a number that does not exist.