Only ChatGPT tags its citation links: 252 of 252 carried utm_source=openai. The other three engines tagged zero of 1,335.

Method, stated before the numbers

Same instrument as the rest of the citation-economy series: 12 pinned buyer questions × 4 engines (ChatGPT, Perplexity, Gemini, Claude) × 3 samples each = 144 answers in one run (2026-08-03), citations read from each engine's returned citation payload. The new cut is mechanical: take all 1,587 cited URLs, parse every query string, and classify each parameter as engine-appended tracking (a parameter the engine adds to mark the click, like `utm_source=openai`) or native to the cited page (a parameter that is part of the page's own address, like a YouTube video id). Every parameter key we found is listed below — nothing is summarized away.

The census: one engine tags everything, three tag nothing

The full parameter census, so the classification is checkable: ChatGPT's 252 URLs carry `utm_source=openai` (252×) plus two parameters native to cited pages (`msockid` 1×, `hl` 1×). Perplexity's 49 param-bearing URLs are all native page addresses: YouTube video ids (`v`, 45×), YouTube locale (`vl`, 4×), a site's own `domain` filter (3×) and one `hs_amp`. Gemini and Claude returned zero query strings across 618 URLs — they cite clean canonical URLs. ChatGPT tagged all 150 distinct sites it cited; the tag is unconditional.

Why this matters: three of four engines send dark traffic

A `utm_source` parameter travels inside the URL, so it survives everything that kills referrer headers — app-to-browser handoffs, privacy strippers, cross-origin policies. A ChatGPT citation click is therefore attributable in any analytics tool, forever. A Gemini, Claude or Perplexity citation click carries no mark at all: unless the referrer header happens to arrive intact, it lands as direct traffic. The practical asymmetry: AI-referral numbers built on URL tags are a ChatGPT-only floor. Our own instrument is a live example — every AI referral we have ever detected on this site is ChatGPT, 4 of 4, and today's census says the other three engines' clicks would have had to survive on referrer headers alone to be counted at all. Some share of everyone's "direct" traffic is AI answers nobody can see.

Limits

• This measures API citation payloads, not consumer UIs. The URLs are what each engine's citation array returns on its API path — the same path our weekly instrument runs. Consumer apps can rewrite links differently (and separately from this measurement, referrer-based detection of the consumer products still works when the header survives).
• One run, one day, one panel. 12 buyer questions about AI visibility for small businesses, 2026-08-03. Tagging behaviour is a vendor default and could change with one deploy — which is exactly why the census is dated.
• Native-vs-appended classification is by inspection. The full key list is published above; with 7 distinct keys across 1,587 URLs, every call is checkable in seconds.
• We have a commercial interest. We sell AI-visibility audits. The dated 144-answer citation cut behind every number is shared with anyone who asks.

The practical reading: segment `utm_source=openai` in your analytics today and you get a clean, durable count of ChatGPT-referred visitors — then treat that number as the floor, not the total, because the other three engines send their clicks unmarked. If you want to know whether the engines are citing you at all — the step before any click — run the four-engine, three-sample measurement on your business in 60 seconds.