Skip to content

AI agents and attribution: how agent visits show up in GA4

GA4 files clicks from AI answers under AI Assistant when the referrer matches a list. It documents no channel for agents that fetch or browse for a shopper: known bots are dropped without a count, and a visit with no referrer or tags looks like Direct.

By , Founder & CEOPublished 6 min read

GA4 has a channel for people who click a link in an AI answer and no documented channel for agents that browse or buy on their behalf. Google's release notes dated 13 May 2026 add an AI Assistant channel, assigned when the referrer matches a list of AI assistants. A visit with no referrer and no tags has nothing to match, and known bots are excluded without a count. Operators publish names, IP ranges and, for some agents, signatures, so your logs can see visits GA4 cannot.

Which agents visit a store, and what do their operators call them?

Four operators document agents that act on a person's request instead of crawling on a schedule, as their own pages read today.

OperatorName in its documentationWhat the page saysHow to verify it
OpenAIChatGPT-Userused "for certain user actions in ChatGPT and Custom GPTs"published IP list
OpenAIChatGPT Work's Cloud browsersigns outbound requests with Web Bot AuthSignature-Agent header set to "https://chatgpt.com"
AnthropicClaude-Userused when "individuals ask questions to Claude"IP list Anthropic links to
PerplexityPerplexity-User"supports user actions within Perplexity"published IP list
GoogleGoogle-Agentagents that "navigate the web and perform actions upon user request"published IP ranges; Web Bot Auth being tested

Google warns that a user agent string can be spoofed, so match the name and the source address. OpenAI adds that its Cloud browser, at launch, "cannot sign in to websites or complete payments": it can browse for a shopper, but another route has to take the money, as the agentic commerce post covers.

What does GA4 show for each kind of visit?

The three cases differ, and only the first has a documented GA4 channel.

  1. A person clicks a link in an AI answer. GA4 files it under AI Assistant if the referrer is on its list, a real improvement for the referral half. Google's AI Overviews and AI Mode are not in that channel; Google counts them in Organic Search. OpenAI says ChatGPT adds utm_source=chatgpt.com to referral URLs from its search results. Ethercycle, a Shopify agency, reports that across ten stores sessions carrying that tag outnumbered sessions with a chatgpt.com referrer 3 to 1, and puts the surplus down to apps that send no referrer (published by ethercycle, not independently audited). The AI-referral posts below cover this case.
  2. An assistant fetches your page to answer a question. ChatGPT-User, Claude-User and Perplexity-User do this. GA4 drops visits it recognises as known bots and won't say how many, and its page doesn't list which fetchers it recognises. No operator page above says whether its fetcher runs a site's analytics tag. A number you can't reconcile isn't a measurement, so count these in your request logs.
  3. A browser agent opens your URL. With no referrer and no tags, GA4 has only its Direct definition to use: "a saved link or by entering your URL". If an agent places an order through a protocol and your pages never load, no page tag has anything to fire on. You can send the purchase to GA4 yourself, but Google's Measurement Protocol page says that sending events solely that way means "only partial reporting may be available".

Shopify's rollout from 21 to 23 September 2026 filters identified bot sessions out of session-related reports by default. Its page calls the classification conservative and says some sophisticated automated traffic "might not be immediately classified", so Shopify's session counts and GA4's can disagree for bot-like visits.

How big is agent traffic, according to the studies?

Cloudflare reports, in a blog post dated 30 September 2026, that daily requests from AI agents on its network grew by more than 1,700% over the past year (vendor-published: Cloudflare offers the bot controls the post describes; not independently audited). Ethercycle reports that AI assistants sent 0.12% of revenue across ten US stores in 2026 through August by Shopify's last non-direct click, and that 25% of Direct sessions were bots against 4% for Google by Shopify's classification. The ten stores are its own retainer clients, and it says ten stores is not a sample for significance tests. Neither source measures agent-placed orders that GA4 did not credit; no public measurement of that is cited here.

How do you check your own store this week?

Build one table with a source and a date range beside every number:

  1. GA4. Sessions and purchases where Session default channel group is AI Assistant. Then list Session source / medium for sources containing chatgpt: the docs don't say which channel a tagged session with no referrer gets, so look.
  2. Shopify analytics. Add the Human or bot session dimension to a sessions report, and don't compare session counts across the September rollout.
  3. Your order system. On Shopify, Sales channels, Agentic shows orders and online store sessions for each AI channel. Elsewhere, ask your commerce or payment provider which order events mark an agent channel.
  4. Logs or CDN. Requests whose user agent contains ChatGPT-User, Claude-User, Perplexity-User or Google-Agent, checked against the operator's IP list or signature.
  5. Reconcile. Agent-channel orders minus GA4 purchases credited to AI Assistant, same dates and time zone. For illustration: 120 orders minus 45 credited purchases leaves 75 orders with no AI Assistant credit in GA4.

Pass: every row has a number, a source and a date range. Fail: a row you can't fill. Write "no data" rather than a guess. The last gap is an undercount for your store, not an industry figure, so keep it as its own line, not inside Direct.

Sources, 30 September 2026: Default channel group (GA4 Help, 2026); What's new in Google Analytics (GA4 Help, 2026); Known bot-traffic exclusion (GA4 Help, 2026); Measurement Protocol overview (Google, 2026); Overview of OpenAI Crawlers (OpenAI, 2026); ChatGPT Work's Cloud browser allowlisting (OpenAI, 2026); Publishers and Developers FAQ (OpenAI, 2026); Google user-triggered fetchers (Google, 2026); Anthropic crawler article (Anthropic, 2026); Perplexity Crawlers (Perplexity, 2026); Bot filtering in Shopify analytics (Shopify, 2026); State of Ecommerce 2026 (ethercycle, 2026); The Internet has a second audience (Cloudflare, 2026).

Frequently asked questions

  • Does GA4 track visits from ChatGPT, Claude or Perplexity agents?
    Only partly. GA4 files clicks from AI answers under AI Assistant when the referrer matches its list. Fetches by the assistants' agents are not reported separately, and GA4 excludes known bots without saying how many. Use your request logs for those.
  • How can I see AI agent traffic in my server logs?
    Filter requests by the user-agent names the operators publish, such as ChatGPT-User, Claude-User, Perplexity-User and Google-Agent. Then check the source IP against each operator's published ranges or its signature, because Google warns that a user agent string can be spoofed.
  • Will an order placed by an AI agent appear in GA4?
    Only if your system sends it. A purchase completed without loading your pages fires no page tag, and Google says events sent only through the Measurement Protocol may give only partial reporting. Compare your platform's AI-channel orders with GA4 purchases for the same dates.

Go deeper: Causal attribution, explained.

Sixty-second versions of these ideas: Causality Engine on YouTube Shorts.

Keep reading

Terms in this article

Browse the full glossary

Your platforms guess.
We run the math.

Upload a GA4 export and see what each channel caused, next to last-click, in 1–2 minutes. The read is yours to keep.

Free, in your browser: your file is not uploaded. The full read is €99, refundable within 30 days. Prices exclude VAT.
Or book a 30-min call.