Where the referrer gets lost
- Native apps: links opened from the ChatGPT mobile/desktop app frequently carry no referrer at all — the visit looks identical to someone typing your URL.
- Copy-paste behavior: assistants present URLs users copy into a new tab — definitionally Direct.
- Referrer policies: strict policies on the assistant's side can strip the header even from web clicks.
- In-app browsers and redirects can each drop the header along the way.
The result: teams look at their referrer report, see a small "chatgpt.com" row, and conclude AI doesn't matter yet — while a meaningful slice of their growing "Direct" bucket is actually AI-driven discovery they can't see.
What each assistant actually sends — measured
On 24 August 2026 we clicked through to our own site from inside each assistant — ChatGPT, Claude, Perplexity and Gemini, on desktop and on an iPhone — and read the arriving requests back from production. Eight surfaces, one click each. This is one day's measurement on one site, so treat it as a snapshot, but it is first-party and dated where most claims in this space are neither.
- Desktop, referrer sent: ChatGPT sends chatgpt.com, Perplexity sends www.perplexity.ai, Gemini sends gemini.google.com. Three of four identify themselves.
- Desktop, nothing sent: Claude sends no referrer at all — the visit is indistinguishable from someone typing your URL.
- iPhone apps, referrer stripped every time: the in-app browsers sent no referrer on any of the four assistants.
- iPhone apps, the tag that saves two of them: ChatGPT appends utm_source=chatgpt.com to the link itself and Perplexity appends utm_source=perplexity — so a stack that reads the link tag still attributes them.
- iPhone apps, unattributable: Claude and Gemini add nothing on iOS. Those visits land in Direct and stay there.
Two practical consequences. First: if your analytics reads only referrers, your mobile AI traffic is invisible — the tag, not the referrer, is how assistant apps identify themselves on iOS. Second: some AI traffic is genuinely unattributable, and a tool claiming to measure Claude's iOS clicks precisely is estimating. We say so instead.
How to see the real number
You can't fully recover a stripped referrer — anyone claiming 100% AI attribution is overclaiming. What an honest measurement stack does is layer three signals:
- Confirmed AI visits: referrers that unambiguously identify an assistant (chatgpt.com, perplexity.ai, gemini.google.com, and the app-specific android-app:// identifiers most tools miss) — plus the utm_source tags the ChatGPT and Perplexity apps write onto their own links, which is how mobile assistant traffic is identified at all. A referrer is what the browser says; a utm parameter is text anyone can write into a URL, so the two deserve different trust levels, and Promvia records which one identified each visit.
- Likely AI visits: direct visits to deep pages that AI engines are known to cite for your tracked questions — labeled separately as inferred, never silently added to the confirmed count.
- AI crawler activity, labelled by purpose: a search-crawler visit (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot) confirms that crawler reached the page — Googlebot's in particular is Search access, which is what AI Overviews runs on, not proof of a fetch for one answer; a user-triggered visit means someone asked about you just now; a training-crawler visit (GPTBot, ClaudeBot, CCBot) is a different thing entirely. All of it lands server-side, before any human clicks.
That confirmed-versus-likely split is a design principle, not a technicality: mixing inferred visits into a confirmed number quietly inflates it, and you end up making decisions on fiction. Promvia reports the two tiers separately for exactly that reason.
A five-minute sanity check you can run today
- Open your analytics' traffic-sources report and search for chatgpt, perplexity, gemini, copilot — that's your confirmed floor.
- Compare your Direct traffic today to a year ago; growth without a matching brand-marketing push is a hint something (often AI) is feeding it.
- Ask ChatGPT and Perplexity your top three buying questions; if they cite your deep pages, some of your Direct visits to those pages almost certainly came from answers.
- Or paste your traffic-source list into our free AI Traffic Checker — it flags the AI rows and totals the share for you, no signup.
Frequently asked questions
Can Google Analytics be configured to catch this?
You can create channel groups for the labeled AI referrers — worth doing — but no GA configuration recovers visits whose referrer was never sent. That share stays in Direct without content-level inference.
How big is the invisible share?
It varies by audience and can't be known exactly — that's the point. Sites we've measured see meaningful gaps between labeled AI visits and inferred-likely visits, but any tool quoting a universal percentage is guessing.
Do UTM links fix this?
You can't put UTMs on links AI assistants generate — but two assistants put them there themselves: in our August 2026 measurement, the ChatGPT and Perplexity iOS apps appended their own utm_source tag to outbound links while stripping the referrer. So reading utm_source recovers their mobile traffic. Treat it as a weaker signal than a referrer, though — a utm parameter is text anyone can write into a URL.