Analytics for AI agents: measuring a web where bots outnumber humans

Microsoft Clarity’s product team published a piece in late 2025 framing the next five years of web analytics around a single shift: the people visiting your site are increasingly not people. They’re agents — autonomous programs that navigate, click, fill forms, and complete checkouts on behalf of a human who is somewhere else entirely.

That framing isn’t marketing. The numbers behind it are blunt. HUMAN Security’s 2026 State of AI Traffic report found AI-driven monthly traffic volumes grew 187% from January to December 2025, and traffic from agentic systems — bots that can navigate authenticated sessions and complete transactions — grew 7,851% year over year. Automated traffic is now growing eight times faster than human traffic. Shopify reports AI-attributed orders are up 11x since January 2025.

If your analytics stack still treats every visit as a human session, it’s missing most of the traffic shift the numbers are describing.

Why traditional analytics break on agents

The architectural problem is older than most people realize. Google Analytics, Matomo, Mixpanel, Amplitude — they all measure the same way: a JavaScript snippet runs in a browser, fires events, and ships them to a collector. That model assumes a renderer.

AI agents mostly don’t render. There are three categories, and only one of them shows up in your dashboards:

Crawlers like GPTBot, ClaudeBot, OAI-SearchBot, and PerplexityBot make raw HTTP requests to harvest content for training and retrieval. They never execute JavaScript. Your GA4 has no idea they were here. They show up only in server logs.

Referral clicks happen when a human reads an AI answer that cites you and follows the link. These do hit GA4 — but they arrive with a referrer of chat.openai.com or perplexity.ai, get bucketed into “Direct” or a generic “Referral” channel, and lose all context about what question prompted the visit. Without a custom channel group, you cannot tell ChatGPT-sourced revenue from random direct traffic.

Agentic browsers are the new and weird category. ChatGPT Agent, OpenAI’s Operator, Perplexity’s Comet, Claude’s computer use — these run real browser sessions on behalf of a user. They execute JavaScript. They fire your analytics events. But the “session” isn’t a person browsing; it’s a model deciding what to click next based on a goal. Your bounce rate, time-on-page, and scroll depth metrics become noise. The agent doesn’t scroll because it doesn’t read like a human. It jumps to the data it needs.

OpenAI’s bots alone account for roughly 69% of all observed AI-driven traffic, with Meta-ExternalAgent at 16% and Anthropic at 11% per HUMAN’s 2026 State of AI Traffic report. If you can’t see them, you can’t see two-thirds of the AI web.

What you should actually be measuring

The instinct is to bolt agent detection onto the existing dashboard and call it done. That’s a half-step. The deeper move is to recognize that agents and humans answer different questions, so they need different metrics.

Here’s the split Lumio recommends to brands and content publishers:

For crawlers (server-log territory)

The question isn’t “did they engage.” It’s “did they get what they came for, and which ones came back.”

• Crawl coverage by bot. Which URLs did GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and Meta-ExternalAgent fetch in the last 30 days? Gaps here mean parts of your catalog are invisible to the model that’s about to answer a shopper’s question about your product.
• Response code mix per bot. A spike in 4xx or 5xx for GPTBot is a silent invisibility event. You will not feel it in revenue for weeks.
• Recrawl cadence. How often does each crawler return? If your prices change weekly and PerplexityBot only recrawls quarterly, the answer it gives next Tuesday is wrong.
• robots.txt and llms.txt compliance. Are the bots honoring what you’ve published? Most do. Some don’t.

None of this is in GA4. All of it is in your CDN or origin logs. Cloudflare, Fastly, and Vercel all expose bot identification on the edge — that’s where this analysis lives.

For referral traffic (the GA4 layer, fixed)

Set up a custom channel group in GA4 that explicitly carves out AI sources. The pattern is well documented — kpplaybook has a clean walkthrough — and it takes about ten minutes. Match referrers and source/medium combinations from chatgpt.com, perplexity.ai, gemini.google.com, claude.ai, copilot.microsoft.com, and the long tail of smaller assistants.

Once they’re carved out, the metrics that matter are not the ones GA4 leads with:

• Conversion rate by AI source. The honest finding from the data is that this varies wildly by category. Some studies show LLM-referred visitors converting several times higher than generic search; others — including a recent analysis of 973 ecommerce sites and 164M transactions — find them converting worse. The split tracks intent compression: high-intent B2B and SaaS see lift, general consumer ecommerce often doesn’t. Either way, blending them into your overall conversion rate hides the signal.
• Landing page distribution. Which pages are AI assistants citing? That’s your live, real-time map of which content the models think is authoritative. It’s also the closest thing to a “ranking” signal you’ll get in the GEO era.
• Revenue per AI-referred session. The number that justifies (or kills) any GEO investment.

For agentic browsers (the genuinely new layer)

This is where most analytics tools have nothing useful to say yet. An agent session looks like a human session in JavaScript, but the behavior signals are inverted. You need to detect agent sessions and analyze them separately, or your aggregate engagement metrics will steadily degrade as agentic traffic grows.

A few practical signals:

• User-agent strings and client hints. ChatGPT Agent, Operator, and Comet identify themselves — for now. Detect them at the edge and tag the session.
• Behavioral fingerprints. Agents tend to load a page, query the DOM directly, and act. No mouse movement, no scroll, sub-second decision times. Clarity, FullStory, and similar session-replay tools can surface these patterns if you ask.
• Form completion rate, not form abandonment rate. For an agent, a half-filled form isn’t hesitation — it’s a missing field in your structured data. If agents are abandoning checkout, the diagnosis is almost never UX. It’s that the agent couldn’t find a price, a shipping option, or a SKU identifier in machine-readable form.
• Task completion rate. Did the agent finish what it came to do? This is the metric Pendo, GoodData, and the new “agent analytics” vendors are converging on, and it’s the one that maps cleanest to revenue.

The attribution problem nobody has solved

Here’s the uncomfortable part. Even with all of the above wired up, you still can’t fully answer the question every CMO is about to ask: how much revenue did AI agents drive last quarter?

The reason is that the agent layer fragments attribution in ways the existing martech stack wasn’t built for. A shopper might:

1. Ask ChatGPT for product recommendations (no visit to your site).
2. Click a Perplexity citation a day later (a referral session, tagged).
3. Have ChatGPT Agent complete the checkout the following week (an agentic session, on a different device, with no cookie continuity).

GA4 sees three unrelated visitors. The actual customer journey is one person, mediated by two different AI surfaces, with the purchase technically performed by a third agent. Last-click attribution gives 100% of the credit to whichever surface the agent happened to complete the transaction from — usually direct, often wrong.

The vendors trying to solve this — Adobe’s LLM Optimizer, Salespeak, Writesonic, SE Ranking, and a wave of new entrants — are mostly stitching server-log bot detection together with GA4 referrer carve-outs and calling the result “agent analytics.” It’s a useful first pass. It is not attribution. Real attribution will require the AI surfaces themselves to expose the conversational context that led to a citation, and there’s no commercial incentive for OpenAI or Google to do that yet.

In the meantime, the honest framing is this: agent traffic is measurable, conversion downstream of agent referrals is measurable, and the gap in your content that agents are skipping is measurable. A single line from “user prompt” to “purchase” is not measurable yet. Anyone selling you that line today is selling you a model output, not a measurement.

Where to start this week

If your organization has done nothing yet, the order of operations is small and concrete:

1. Pull 30 days of server logs and group by user-agent. Count GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Google-Extended, Meta-ExternalAgent, Bytespider, and Amazonbot separately. You will be surprised — usually unpleasantly — at the volume.
2. Add an AI custom channel group in GA4. Ten minutes. It will immediately split out a chunk of “Direct” traffic that isn’t direct.
3. Audit which of your top 50 pages return 200s to crawlers. Any that don’t are silently invisible to the AI surface that matters most for that topic.
4. Tag agentic browser sessions at the edge. Cloudflare and Fastly both let you do this with a few lines of config. Pipe the tag into your analytics as a custom dimension.
5. Stop optimizing aggregate engagement metrics until you’ve separated agents from humans. Time-on-page, bounce rate, and scroll depth are now blended numbers across two populations that behave nothing alike. Optimizing the blend will move you in the wrong direction for at least one of them.

The five-year arc Microsoft is describing is real, and the underlying numbers from HUMAN, Shopify, and the bot-detection vendors all point the same direction. The web is becoming a place where most of the visitors aren’t people, the ones that aren’t people are increasingly the ones driving revenue, and the analytics tools every team relies on were designed for a world that no longer exists.

The fix isn’t a new dashboard. It’s deciding, deliberately, what you want to know about a population your old tools were never built to see.

---

If you want help auditing how visible your catalog is to crawlers, how your AI referral traffic is converting, or how to instrument agentic browser sessions before they become a majority of your traffic, get in touch. It’s the kind of work Lumio does every day.