How I track AI agent traffic on our website

Key takeaways

• "AI traffic" is three different things: humans referred by AI chatbots, AI crawlers, and AI agents that browse and buy for people. Each lives in a different data source.

• Agents that act like shoppers land inside your funnels as ordinary sessions, so you can't see how they experience your site differently from people.

• Fullstory tracks AI agent traffic from count to cause: how many agents visit, where they go, and what each one was trying to do.

  ---

If you only look at AI referrals for "AI traffic", you're looking at an incredibly small slice. AI chatbot referrals are under 1% of traffic on our site.

Agent-like and automated traffic, on the other hand, is about a third of our sessions. So let's be honest with ourselves, these chat tools were never built to send clicks (apparently neither was Google–only half kidding for my SEO friends). The visits, the action, it all comes after the education in those platforms.

The bigger story for agents actually sits in your Direct, paid and all the other channels, where sessions carry agent fingerprints. Odd screen resolutions like 800 x 600 are a start. Behavior no person would produce like no curved mouse movements or none at all may signify something closer. Lots of people have been hunting these patterns in GA for years, but Fullstory picks up signals that GA4 and other tools just don't.

Don't forget the human behind the agent either. Someone asked a question and sent that bot to answer it. When an agent reads our pricing page, a person almost certainly wanted pricing. The goal didn't change. The visitor did.

Automated traffic now makes up more than half of all web traffic, according to Cloudflare. Some of it is AI agents acting for people: booking flights, comparing products, using software, filling out forms, completing checkouts. Those visits can land in your analytics as human sessions. And Seer Interactive tested OpenAI's Operator agent, its visits showed up in GA4 as Bing organic traffic. That's muddied attribution if I've ever seen it.

Customers tell us the same thing: they need to identify and separate agent and bot traffic from GA4 so year-over-year results stay reliable, and they need to tell helpful agents from malicious bots, especially in financial services.

So, here's how I track AI traffic.

Step 1: Understand the three kinds of AI traffic

• AI referral traffic is people clicking a link in ChatGPT, Perplexity, Gemini, Claude or Copilot. These are humans, and they belong in your main analytics as Referral traffic.

• AI crawlers and fetchers are bots like GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot. Some collect content to train models, some fetch pages to answer a user's question, others are search crawlers. They show up in server and CDN logs, and many never execute JavaScript.

• AI agents are the blind spot. Computer use agents like Meta's Muse open pages, click, fill forms and check out on a person's behalf, and their sessions can look like human visits inside your funnels.

Mix them up and you might get bad decisions. Agent sessions blend into your numbers, so you can't see how agents and people experience your site differently.

Here's what the filters for two simple segments look like:

LLM-Referred Traffic

Agentic Traffic

Step 2: Track AI referral traffic in your tool of choice

GA4 is still a great home for people sent by AI tools. Google Analytics now groups recognized AI chatbot referrals into an "AI Assistant" channel in the default channel group, with ChatGPT, Gemini and Claude among the recognized sources. No setup is required. Installing Fullstory is just as simple, and it gets data flowing.

Two gaps to cover in GA4:

• Unrecognized sources. Google hasn't published the full referrer list. Build a custom channel group with a regex rule on session source for any AI tools it misses, such as Perplexity or Copilot. You can do the same with many more options in Fullstory segments.

• Missing referrers. Traffic from AI apps and in-app browsers often arrives without a referrer and lands in Direct. If there's no referrer data, there's not much any tool can do. Behavioral signals can at least sort those visits into another bucket.

GA4 tells you how many people AI tools sent and what they converted on. Fullstory sees the same and what those visitors did on the page: where they hesitated, what they skipped and what got them to convert.

Step 3: Use server logs, CDNs and bot managers for crawlers

Server and CDN logs see every request, including bots that never run JavaScript, which makes them the right source for AI crawler monitoring.

Filter user agent strings. Declared crawlers identify themselves. OpenAI, for example, documents three separate user agents: GPTBot for model training, OAI-SearchBot for ChatGPT search and ChatGPT-User for pages fetched on a user's request. Filter your access logs for these and their equivalents from other AI companies, and keep unknown automated clients in their own bucket instead of guessing.

Use your bot manager's scores. Bot management tools score good vs. bad bots. Some enterprise teams want those CDN bot scores in their analytics so they can tell helpful crawlers from aggressive scrapers.

Logs show volume and paths. They don't show intent, whether a purchase was completed, or where an agent got stuck. I check crawlers in our Netlify logs, separate from everything below.

Step 4: Segment agent sessions

Agents that use a real browser run your tags like a person would, so header checks and IP lists miss many of them. They also change quickly (see Step 7).

Start with behavior, not referrers. Odd screen sizes like 800 x 600, unusual resolutions and non-human click patterns are strong tells, and they show up whether the visit was tagged Direct, paid or organic.

In Fullstory, Suspicious Activity signals (available on Business and Enterprise plans) let you filter sessions to AI Agent. Save that as a segment, then compare agent sessions to human ones side by side. I look at where each goes, what each clicks, and where each gets stuck. The gap between them is a user experience read.

I didn't build the segments in Step 1 by hand. I connected Claude to Fullstory through the MCP and had a conversation. I started with this prompt:

Create a Fullstory segment for AI traffic:

• Referral traffic - from all known referrals including but not limited to ChatGPT, Gemini, Claude, etc. Where the referral hostname or referring URL matches these common sites.

• AI agents - using suspicious activity, mouse movements, fast typing or auto fill, known AI agents, crawlers, assistants, etc. Include common or odd screen resolutions that may signal CUAs and other types of agentic traffic.

Use other helpful filters I may have missed in these definitions with what Fullstory can track.

When created, output in a line chart for each so I can see trends over time. Give me a narrative with brevity on trends, top pages, user behaviors, and common patterns, and potential areas of optimization through StoryAI.

Step 5: See where agents go and what they complete

Once a clear segment is defined, the question shifts from "how many?" to "what are they doing?" Here's what we see on our own site.

On www.fullstory.com:

• Referred AI chat sessions ran roughly 40 to 70 per week through September and October. That is shockingly low if you just follow LinkedIn hype. It’s still likely an undercount, because some of that traffic arrives without a referrer we can see.

• Sessions with agent-like or automated behavior ran 6,000 to 12,000 a week. That segment is broad on purpose: it includes crawlers, headless browsers and other automated sessions, not only AI agents. That’s about 175 for every human an AI chat sends us, and the number is up about 40% since last November. They’re reading the same pages buyers do, which means a growing share of a site’s audience never shows up in funnels if they're not set up properly.

The most-visited pages in those sessions were our homepage, then our pricing page, plus our careers page, one of our paid landing pages, and a few others. Agents go where buyers evaluate.

And agents don't always come in through the front door. On one travel site, Meta's Muse landed on flight results more often than the homepage: it worked out the URLs and skipped straight to inventory. Muse sessions also spend much more time inactive than human sessions, a reminder that agent journeys look different from human ones even when they end in the same purchase.

I sent my Muse to our website

I call him Pip. He was instructed to browse our site, and Fullstory picked it up:

• I told Pip to go to https://www.fullstory.com/?hi-im-muse (the query string let me track exactly where he was)

• I told him to visit our demo page and type his name in the first field to verify it was him.

• I took over his browser from inside the app. Below, you can see the screen change from a mobile view to a desktop view the moment I did.

The recording shows all of it, including my takeover and the mobile-to-desktop switch. At the very bottom, you can see every page visited and a feed of its behavior.

To see this on your site, you can use the Meta Muse activity dashboard template, which shows commerce (orders, revenue, average order value), engagement, human takeovers, errors and navigation. Duplicate it and swap in an AI Agent segment to cover a whole class of agents instead of one.

Okay enough of the Muse fun, back to the post.

Step 6: Decide whether to block, help or watch

In our conversations with brands about agents like Muse, teams land in three broad camps: block it, welcome it, or not sure yet.

• Block when agent behavior looks like fraud or probing, such as repeated form attempts or credential testing. Enforce it at the CDN, WAF or application layer.

• Help when agents are completing real purchases or research for your customers. Find the broken controls and steps that force a human to take over, and fix them.

• Watch when you're still sizing the impact. Keep measuring until the data shows revenue, risk or both.

My take: let agents explore, with guardrails. Seeing which elements agents read and what information they pull is useful.

Whichever camp you're in, find out what an agent came to do before you block it.

Step 7: Keep monitoring as agents change

Around October 16, Muse traffic in our customers' segments dropped sharply. People didn't stop using Muse. Muse changed how it behaves. We're playing cat and mouse with it right now, and honestly it's a fun time to work on this. We rebuilt our detection in hours from sessions we had already captured, with no new tracking, because Fullcapture had already logged every interaction. Header checks and IP lists can't adapt that fast.

Review your agent segments and how they compare to human sessions regularly, and weekly around peak sales periods, major AI launches, or search updates.

Remember, your next buyer might not be a person

Some of the visitors on your site this quarter won't be people. They'll arrive with a task, skip your homepage, go straight to pricing or inventory, and either finish the job or quietly fail. GA4 will count them. Fullstory will show you what they came to do, and whether your site let them.

Referrals only tell you where a visit came from. Fullcapture records everything that happens on your site, and you can ask it questions through MCP to get the intent behind the user or the agent.

Ready to see your own agentic traffic? Start with a demo, or get started with Fullstory for free.