The exact channel-group setup that surfaces ChatGPT, Claude, Perplexity, Gemini, and Copilot referrals. Plus the four things GA4 will still miss when an agent is doing the shopping.
The setup is ten minutes. The interesting part is what it doesn't catch — and the diagnostic gap that opens up once your team starts asking why agent-driven sessions don't convert.
Each of these services sends a real `Referer` header when a user clicks a citation in the answer panel. By default GA4 treats most of them as Direct because they aren't in the canonical Referral list. Adding them as a custom channel group exposes them.
| Service | Referrer hostnames | Note |
|---|---|---|
| ChatGPT | chatgpt.com, chat.openai.com | Sends referrer on user-driven citation click. GA4 sees it as Direct without a rule. |
| Claude | claude.ai, claude.com | Citation click from Claude's answer surface. |
| Perplexity | perplexity.ai, www.perplexity.ai | Source-link click from a Perplexity answer. |
| Gemini | gemini.google.com | Direct citation click from Gemini's answer panel. Google AI Mode citations appear under google.com. |
| Copilot | copilot.microsoft.com | Microsoft Copilot citation click. |
This adds an "AI Assistants" row to your acquisition reports. Existing channel groups are not affected.
GA4 Admin → Data display → Channel groups → Create new channel group. Name it "AI Assistants."
Add a new channel inside the group. Match condition: Source matches regex.
The regex below covers the five services above plus the new browser-mode surfaces (ChatGPT Atlas, Perplexity Comet) where they pass referrers.
Reports → Acquisition → User acquisition → set the channel group dropdown to "AI Assistants" → check that you see traffic. Backfill is up to GA4; expect 24–48h for historical data.
chatgpt|chat\.openai|claude\.ai|claude\.com|perplexity\.ai|gemini\.google|copilot\.microsoft|comet\.perplexityIf you use Google Tag Manager, you can also push the assistant identity into a custom dimension at the dataLayer level — useful for cross-segment funnels. Reach out if you want the GTM snippet.
The setup above catches the slice of AI traffic that travels with a referrer. Most of it does not.
Loamly's 2026 study of 446,000 AI-driven sessions found 70.6% had no referrer header. Headless browser agents (Operator, ChatGPT Atlas, Perplexity Comet, Claude with Playwright MCP) navigate via direct URL fetch. GA4 logs them as Direct, regardless of how many regex rules you add.
When ChatGPT shows a product card inside the chat — image, price, link — and the user buys via the in-chat path, your site sees zero traffic. OpenAI's Instant Checkout flow runs through Stripe and a Shopify connection that bypasses your storefront entirely.
GA4 can tell you the visit happened. It cannot tell you that the agent reached the product page, tried to click a roleless variant selector, retried twice, and exited. The entire failure mode lives between landing-page and conversion-event, in a black box.
Hotjar, Contentsquare, FullStory, Lucky Orange capture mouse movement, scroll, click. Browser agents do none of those. Heatmaps of agent sessions are flat. Recordings are silent.
These are the user-agents you'll see in server logs but not in GA4 (because they don't fire client-side scripts). Sources linked at the bottom of the page.
| Bot | Owner | Purpose | robots.txt directive | Visible in GA4? |
|---|---|---|---|---|
| GPTBot | OpenAI | Training data crawl | User-agent: GPTBot | No — server log only |
| OAI-SearchBot | OpenAI | Search index for ChatGPT search | User-agent: OAI-SearchBot | No — indexer |
| ChatGPT-User | OpenAI | Live fetch on user instruction ("go to X and tell me Y") | User-agent: ChatGPT-User | Sometimes — fires GA if it loads JS |
| ClaudeBot | Anthropic | Training data crawl | User-agent: ClaudeBot | No |
| anthropic-ai | Anthropic | Older training crawler (still in use) | User-agent: anthropic-ai | No |
| Claude-User | Anthropic | Live user-driven fetch via Claude.ai | User-agent: Claude-User | Sometimes |
| PerplexityBot | Perplexity | Index crawl | User-agent: PerplexityBot | No |
| Perplexity-User | Perplexity | Live user-driven fetch | User-agent: Perplexity-User | Sometimes |
| Google-Extended | Training opt-out token (block to exclude from Gemini training) | User-agent: Google-Extended | No | |
| GoogleOther | Multi-purpose Google fetcher | User-agent: GoogleOther | No |
Sources: OpenAI bot documentation, Anthropic bot documentation, Perplexity crawler docs, Google special-case crawlers.
GA4 catches the arrival — the visit happened. Hotjar and Contentsquare capture human behaviour on the page. Cloudflare and HUMAN catch and block bots. Nothing in your stack tells you what an agent did between landing and giving up. That gap is where Serge sits.
Free scan · 30 seconds · no signup
Paste your domain. Serge crawls your store in about thirty seconds — completely deterministic, no AI in the scan loop — and hands back a fast snapshot: where agents may get blocked, what to inspect first, and suggested fixes for your team to verify. No signup. Forward the result to your front-end lead.
The scan is the starting point. The deeper picture comes from session visibility, replay, and briefing.
No. Custom channel groups in GA4 are additive — they sit alongside the Default Channel Group. Reports default to Default; you switch to AI Assistants in the channel-group dropdown when you want the breakdown.
Sometimes it does. Live user-driven fetchers (ChatGPT-User, Claude-User, Perplexity-User) execute on the user's behalf and may run client-side scripts. When they do, gtag.js fires and you'll see the visit. The catch is that the user-agent string contains "ChatGPT-User" — GA4's standard reports treat unfamiliar user-agents as direct. You'd need a custom dimension on user-agent to break those out.
Separate question. GPTBot, ClaudeBot, anthropic-ai, and Google-Extended are training-data crawlers — blocking them removes you from the training set but does not affect agent shopping at runtime. ChatGPT-User and Claude-User are runtime fetchers; blocking those breaks live agent shopping. Most commerce sites want to allow runtime, decide separately on training.
Yes — server-side analytics (Plausible, Fathom server module, raw access logs through a parser) sees the user-agent on every fetch, including agents that don't run JavaScript. This catches GPTBot, OAI-SearchBot, ClaudeBot, etc. It still doesn't tell you whether the agent succeeded — only that it visited. The server log is the floor; the diagnostic is what's missing on top.
Both are full browser modes (Comet launched 2025, Atlas launched 2026). They run real Chromium under the hood and execute JS, so GA4 fires. The user-agent contains a marker ("Atlas" or "Comet") and the referrer is sometimes preserved, sometimes stripped — depends on how the user navigated. Add the relevant fragments to the regex; expect the surfaces to keep changing for the next 12 months.
Treat it carefully. Google AI Mode and AI Overviews show citations that link to your site; the click-through arrives with referrer google.com, indistinguishable from a normal Google search click. GA4 cannot separate AI-Mode-driven google.com clicks from organic-search google.com clicks without help from Search Console's AI-Mode metrics, which roll out unevenly.
Serge uses analytics cookies to understand which pages teams find useful and where the funnel breaks. Essential cookies (login, language) always run. See what each one does →