How ChatGPT, Claude, Perplexity & Gemini Pass Referrers: A Field Guide
Each AI engine handles outbound traffic differently — some pass referrers, some strip them, some auto-append UTMs. The complete reference table of detection patterns for every major AI engine in 2026.
← Back to AI Search Attribution
Every AI engine handles outbound clicks differently. Some pass a referrer header. Some strip it. Some append UTM parameters. Some behave differently on iOS vs Android vs desktop browser.
This is the field guide. It covers every major AI engine, how they pass traffic signals, and which signals are reliable enough to build attribution on.
Why Referrer Behavior Varies
The referrer header is set by the browser — but AI engines influence it through:
- Link rendering method — links inside an
<iframe>may strip referrers via thereferrerpolicyattribute - App vs browser — mobile apps (ChatGPT iOS, Perplexity iOS) often suppress referrers
- Privacy settings — users with aggressive privacy extensions may strip referrers regardless of the engine
Referrer behavior is probabilistic, not deterministic. The table below reflects typical behavior across 10,000+ measured click events.
The Complete Referrer Reference Table
| Engine | Referrer domain | Reliability | Auto UTM | Notes |
|---|---|---|---|---|
| ChatGPT (browser) | chatgpt.com | ~40% | No | Desktop passes more than mobile |
| ChatGPT (iOS app) | chatgpt.com | ~10% | No | App aggressively suppresses |
| Claude (browser) | claude.ai | ~50% | No | More consistent than ChatGPT desktop |
| Claude (iOS) | claude.ai | ~15% | No | App strips referrer |
| Perplexity (browser) | perplexity.ai | ~85% | No | Most reliable of the major engines |
| Perplexity (iOS) | perplexity.ai | ~60% | No | Still passes more than rivals |
| Gemini (browser) | gemini.google.com | ~70% | No | Shows as Google in GA4 without custom rules |
| Gemini (AI Overview) | google.com | ~60% | Sometimes | Indistinguishable from organic in GA4 |
| Copilot | bing.com | ~75% | Sometimes | Microsoft’s Bing attribution logic |
| Meta AI | l.meta.ai | ~30% | No | Aggressive link wrapping |
| You.com | you.com | ~90% | No | Behaves like a standard referrer |
| Grok (xAI) | grok.com | ~50% | No | Patterns still stabilizing |
Engine-by-Engine Detail
ChatGPT
ChatGPT’s referrer behavior tightened significantly in 2024–2025. Desktop browser sessions pass chatgpt.com as referrer on roughly 40% of clicks. The drop to ~10% on iOS is because the ChatGPT iOS app opens links in an in-app browser with strict referrer policies.
Detection patterns:
chatgpt.com
chat.openai.com (legacy, still active)
What you miss: ~60% of ChatGPT desktop sessions and ~90% of iOS sessions arrive with no referrer signal. Without a secondary attribution layer, these are invisible.
Claude
Claude is more consistent than ChatGPT on desktop — approximately 50% of desktop sessions pass the claude.ai referrer. Unique to Claude: when Claude uses Brave Search for web lookups, the referrer may appear as search.brave.com rather than claude.ai. These sessions are Claude-influenced but Brave-attributed.
Detection patterns:
claude.ai
Perplexity
Perplexity is the most attribution-friendly AI engine. ~85% of desktop sessions pass the full perplexity.ai referrer — likely a deliberate product decision, as Perplexity positions itself as citations-first and wants publishers to see their traffic.
The GA4 problem: Despite reliable referrers, GA4 misclassifies Perplexity as “Organic Social” because perplexity.ai matches social heuristics. Fix this with a custom channel group rule before trusting your Perplexity numbers.
Detection patterns:
perplexity.ai
www.perplexity.ai
Gemini
Gemini routes through Google’s infrastructure. Conversational Gemini sessions arrive with gemini.google.com — which GA4 sees as Google and labels Organic Search. Traffic from Google AI Overviews (the inline AI summary box in search results) is even harder to separate — it arrives with google.com as referrer with no reliable distinguishing path.
Detection patterns:
gemini.google.com (conversational — distinguishable)
google.com (AI Overviews — NOT distinguishable from organic without source= param)
Copilot (Microsoft)
Copilot routes through Bing’s infrastructure. Referrers pass as bing.com/chat or bing.com. Microsoft occasionally appends msclkid — similar to gclid on Google — which enables order-level Microsoft attribution when present.
Detection patterns:
bing.com
msclkid= parameter (when present)
form=MA13FV (Bing Chat query parameter)
The Multi-Signal Detection Approach
No single signal is reliable across all engines and platforms. Robust AI attribution requires combining:
| Signal | What it catches | Reliability |
|---|---|---|
| Referrer header | 40–85% of sessions (engine-dependent) | Medium-high |
| UTM parameters | Sessions you’ve pre-tagged | High when present |
Shopify referring_site field | Recorded at order creation, survives ad blockers | Medium |
Shopify landing_site field | First page visited, correlate with your AI-optimized content | Low-medium |
Inxy’s attribution layer combines all four. The result captures roughly 85–90% of AI-originated revenue, versus the 30–40% captured by referrer-only approaches.
Next: Setting Up AI Source Attribution on Shopify — the four implementation options ranked by effort and accuracy.