Guide · Last updated November 7, 2026 · 7 min read
How to measure the AI traffic your analytics misses
Guide
Your analytics sees people who click through from AI answers. It misses the agents themselves, by design. To measure AI traffic, track AI referrals in your analytics, count agent visits in your server or CDN logs, and check which agents are real.
Who this is for
Who should read this.
- Marketing or growth lead
- A GA4 setup for AI referrals and a weekly scorecard
- SEO or content lead
- Which pages agents read, and which ones they send people back to
- Web or IT lead
- Where agent visits show up in logs, and how to tell real from fake
The three kinds of AI traffic
AI traffic comes in three kinds, and standard analytics sees only one of them well.
| Kind | Example | Shows in GA4? | Why |
|---|---|---|---|
| AI referrals | A person clicks a link inside a ChatGPT answer | Yes, as a referral | A human browser loads your page and runs your analytics script |
| Agent visits | ClaudeBot, ChatGPT-User, or OAI-SearchBot reads a page | No | Most agents never run the script, and GA4 removes known bots automatically. Google says you can't turn that off or see how much it removed (Google). |
| Browser agents | An agent drives a real browser to shop or fill a form | Sometimes | It runs your script, so it can look like a human visit (our research) |
The first kind tells you what AI sends you. The second tells you what AI takes. You need both to judge the trade.
Step 1: Track AI referrals in GA4
Since May 13, 2026, GA4 has a built-in AI Assistant channel for people arriving from apps such as ChatGPT, Gemini, Deepseek, Copilot and Grok (Google). Start there. To choose exactly which apps count, give AI a channel of your own. This takes about 15 minutes.
- In GA4, go to Admin, then Data display, then Channel groups.
- Copy the default channel group and name it "Default + AI."
- Add a new channel named "AI assistants."
- Set the condition to Session source, matches regex, and paste the pattern below.
- Drag "AI assistants" above "Referral" so it catches these visits first. Save.
chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|meta\.ai
Add new AI apps to the pattern as they appear in your referral list. Some apps hide the referrer, so part of your AI traffic still lands in Direct. Treat this channel as a minimum count.
Step 2: Count agent visits in your logs
Agent visits live in your server or CDN logs, where every request is recorded with its user agent name. Your hosting or CDN dashboard often has a filter for this.
- Export a week of request logs, or open your CDN's traffic view.
- Filter the user agent field for the names below.
- Count requests per agent, per day, and note the top pages each one reads.
| Kind | User agent names to filter |
|---|---|
| Training crawlers | GPTBot, ClaudeBot, CCBot, Applebot-Extended |
| Search crawlers | OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot |
| Assistants | ChatGPT-User, Claude-User, Perplexity-User |
The glossary lists more agents and what each one does. A name in your logs is only a claim. Step 3 checks it.
Step 3: Separate real agents from fakes
Any scraper can call itself GPTBot. Before you report agent numbers, match each request's address against the operator's published list.
| Operator | Where to check | What it covers |
|---|---|---|
| OpenAI | openai.com/gptbot.json, searchbot.json, chatgpt-user.json | GPTBot, OAI-SearchBot, ChatGPT-User (OpenAI) |
| Anthropic | Anthropic crawler page | ClaudeBot, Claude-SearchBot, Claude-User |
| Others | Each operator's crawler documentation | Varies |
A request that claims a famous name but comes from an address outside that company's list is an unknown scraper. Report it separately. Counting fakes as real agents inflates your AI numbers and hides the scraping. Our research on the verifiability gap shows why: most known bots publish nothing a site can check.
Doing this by hand works for a one-time audit. For ongoing tracking, a tool at your front door, like DNA, checks every request as it arrives.
A weekly AI traffic scorecard
Six numbers, checked weekly, tell you what AI takes from your site and what it sends back.
| Metric | What it tells you | Where to get it |
|---|---|---|
| AI referral sessions | People AI sent to your site | GA4, AI assistants channel |
| AI referral conversions | Sales or sign-ups from those people | GA4, AI assistants channel |
| Verified agent requests, by kind | How much each kind of agent reads | Logs, after step 3 |
| Unverified requests using famous names | How much scraping hides behind real names | Logs, after step 3 |
| Top pages read by assistants | Which pages answer customers' questions | Logs, filtered to assistants |
| Agent requests per AI referral | How much AI reads for each person it sends back | Divide the two numbers above |
The last number is your crawl-to-refer ratio. Watch it by page. A page that agents read often but that sends no one back needs a reason to click: a tool, a price, a next step.
Questions
Why does some AI traffic show up as Direct?
Some AI apps open links without passing a referrer, so GA4 can't tell where the visit came from. Your AI assistants channel shows the minimum, and part of Direct is AI too.
Do I need to track every agent?
No. Start with the ones tied to customers: assistants like ChatGPT-User and Claude-User, plus AI referrals. Add training and search crawlers once those numbers are steady.
Is a high crawl-to-refer ratio bad?
It depends on the page. A help article that answers questions inside AI apps may read often and refer rarely, and still win customers. A product page with the same pattern is losing sales.
How often should I check the scorecard?
Weekly. Agent traffic shifts as AI companies launch and retire agents, and a weekly check catches changes before they show up in revenue.
Sources
- Overview of OpenAI crawlers and published IP lists, OpenAI
- Anthropic crawlers, Anthropic
- Known bot-traffic exclusion, Google Analytics Help
- Default channel group, Google Analytics Help
- How to track AI agents when Google Analytics can’t see them, D3 research, Nov 2026
- The verifiability gap, D3 research, Aug 2026