Agentic browsers are not classic crawlers
Until recently, machine traffic from AI providers was above all one thing: a crawler. Services like GPTBot or ClaudeBot fetch content to train models or ground answers in sources. They identify themselves through a distinct user agent, come from the provider's data centers and can be matched by known names — how to do that is covered in the piece on measuring AI crawler traffic. Agentic browsers work in a fundamentally different way.
An agent browser like Comet or Atlas is not a background crawler but an application that works in sync with the user. The person sets a task ("find the cheapest offer", "book an appointment"), and the browser then navigates real pages on its own, clicking, scrolling and filling fields. From the server's point of view this looks largely like a human session: real interaction, a normal page path, often a mainstream browser core. That is precisely why the old detection patterns no longer hold. An automated request often gives itself away through navigator.webdriver or automation signatures, yet not every agent sets those markers, and the line between "the user's tool" and "an autonomous bot" is blurring.
The old split of "human or bot" no longer suffices. Three classes make sense: the AI crawler that fetches content for training and answers; the agentic browser that acts in real time on behalf of a person; and the real human at the device. Each class deserves its own treatment in reporting — the crawler filtered out, the agent reported as its own channel, the human measured cleanly.
The numbers: how fast agentic traffic is growing
The pace is remarkable. HUMAN Security puts the growth of agentic traffic since early 2025 at roughly 6,900 percent (HUMAN Security). In its April 2026 analysis, agent browsers already make up almost three quarters of all agentic traffic — around 71 percent (HUMAN Security). Within those agent browsers, Comet leads with about 48 percent, followed by Atlas with about 21 percent (HUMAN Security). What was once a fringe phenomenon has become a measurable channel within a year.
In parallel, the overall picture of web traffic has shifted: automated requests now account for 51 percent of all web traffic for the first time, overtaking the human share, with 37 percentage points coming from bad bots (Imperva). The rise is attributed explicitly to the broad availability of AI tools (Imperva). And agentic traffic is not just volume: in retail, AI-referred traffic to US sites rose 138 percent year over year in May 2026 (Adobe Analytics), AI traffic to US retailers jumped 393 percent in the first quarter of 2026 (Adobe Analytics), and since October 2024 the growth adds up to 1,324 percent (Adobe Analytics). Already 39 percent of respondents said they use AI for online shopping (Adobe Analytics).
| Metric | Value | Source |
|---|---|---|
| Agentic traffic growth since early 2025 | +6,900% | HUMAN Security |
| Share of agent browsers in agentic traffic | around 71% | HUMAN Security |
| Comet share among agent browsers | around 48% | HUMAN Security |
| Atlas share among agent browsers | around 21% | HUMAN Security |
| Automated share of web traffic | 51% | Imperva |
| AI-referred retail traffic (May 2026, YoY) | +138% | Adobe Analytics |
These figures do not mean every report is suddenly useless. They do mean a blind spot forms when agentic sessions seep unnamed into the existing channels — sometimes as direct, sometimes as referral, sometimes as an apparent human. And that blind spot is growing at triple-digit rates right now.
Why Atlas appears as direct and Comet as referral
The difference lies in how the link is opened. Atlas typically opens shared or in-conversation links in an internal sandbox or webview context and passes no referrer header. Without a referrer, an analytics tool cannot determine the origin and records the session as direct or as "not set" (MarTech). Comet, by contrast, in many cases passes a referral from perplexity.ai when a result is clicked, so it shows up as a referral in the report (MarTech). Two agents of the same kind — two entirely different traces in your reporting.
The pattern is not new, just larger. A missing referrer has long produced "dark" traffic that is wrongly counted as direct. In practice, one analysis finds around 71 percent of AI traffic lands as direct in classic, cookie-based reporting (Seresa). And direct traffic was never as clean as it looks: studies of so-called dark social show that with active email, social and content marketing, typically 30 to 60 percent of direct traffic actually comes from other channels (SparkToro), while direct traffic overall makes up around 37 percent of all visits (SparkToro). Agentic traffic noticeably amplifies this old problem.
A high direct share is often read as a strong brand — "people type us in directly". In practice it now contains a growing share of sessions with no referrer passed, including agent browsers like Atlas. Crediting direct traffic to the brand unchecked overstates it and simultaneously understates the AI channel.
What the misattribution costs the business
Misattributed traffic is not a cosmetic issue; it distorts decisions. When agentic sessions vanish into direct, a double error forms: the brand looks stronger than it is, and the AI channel looks weaker than it is. Budget then flows into the seemingly strong direct source, while the actually growing AI channel stays underweighted.
Yet that channel is valuable. According to Adobe, AI-referred traffic has recently converted noticeably better than other sources, with revenue per visit around 37 percent above non-AI traffic (Adobe Analytics). Someone arriving via an AI assistant typically spends 48 percent more time on the page, views 13 percent more pages and shows a 12 percent higher engagement rate (Adobe Analytics). This exact high-value traffic is lost when it disappears unnamed into direct. Add the distortion from purely automated bots: they inflate page views and bounce rates and make campaigns look better or worse than they are — why clean data quality matters here is covered in the piece on filtering bot traffic.
Every mis-sorted session shifts budget in the wrong direction. A growing, well-converting AI channel disguised as direct traffic stays invisible — and therefore underfunded. Separating human, agentic and automated sessions creates the basis for decisions that match actual behavior.
Why the first-party context stays accurate
The decisive lever is the measurement context. Much of the distortion arises because signals are masked or never passed in the third-party context. This affects more than the referrer: IP addresses are increasingly obscured in third-party contexts too. How strongly that hits origin analysis is described in the piece on Chrome IP Protection and geodata. In the first-party context, by contrast, your own site sees the request as it arrives — with user agent, a set or missing referrer, and the actual network.
Cookieless first-party measurement reads exactly these request signals instead of relying on device storage or a third-party script. This has two benefits: first, attribution stays more stable because it does not depend on the referrer policy of someone else's context. Second, measurement runs without touching browser storage, which typically makes it independent of the cookie consent requirement — an advantage that matters again now that the third-party cookie plans have ended, as the pieces Privacy Sandbox is dead and the analysis of Consent or Pay show. Origin can additionally be classified via a native IP geo module that distinguishes data centers from consumer networks.
Spotting agentic sessions: the signals
Robust classification relies not on a single trait but on a bundle of signals from the request context. Four groups carry most of the separating power:
Agent browsers often carry distinct identifiers such as Comet or Atlas in the user agent or client hints. These markers are the most direct clue — provided the measurement sees the untampered string, which is usually the case in the first-party context.
Many automated requests set navigator.webdriver to true or show other automation signatures. Not every agent does, but where the marker is present it gives a clear signal — and XICTRAQ already blocks automated browsers at this point in standard tracking.
A session coming from a data center rather than an end-user network is a strong bot signal. An ASN-based classification recognizes hosting networks without wrongly catching mobile carriers — the basis for an accurate data-center filter.
Machine behavior stands out through patterns: unnaturally even click intervals, missing mouse micro-movements, no access to device storage. Individually each signal is weak; combined they form a robust picture.
How XICTRAQ separates human from agentic sessions
XICTRAQ measures in the cookieless first-party context and classifies every session from the signals above. An ASN-based data-center filter discards requests from hosting networks without catching real mobile users. Known agent identifiers such as Comet and Atlas are recognized and carried as their own type, rather than letting them vanish silently into direct or referral. The native IP geo module classifies origin and separates data centers from consumer networks.
The order matters: purely automated bots belong out of the metrics, so page views, dwell time and bounce rate reflect the real audience. Agentic sessions, by contrast, are not deleted but made visible as their own channel — because, as the Adobe figures show, they represent growing and well-converting traffic (Adobe Analytics). The result is reporting that measures the human cleanly, sorts out the bot and names the agent.
The classification works without accessing device storage and relies solely on the request context. That keeps measurement viable under current privacy requirements and typically independent of the consent banner — a difference that protects reach and data quality at the same time in everyday use.
Attribution: making the AI channel visible
Separation is half the job, attribution the other. Once agentic sessions exist as their own type, the AI channel can be measured instead of lost in direct. This is closely related to how visible a brand is in AI answers at all — the piece on measuring AI visibility and GEO referral traffic goes deeper. It is also worth watching distorting sources: how to keep artificial referrer and ghost traffic out of your reporting is shown in the piece on referrer spam and ghost traffic.
In practice that means: Comet referrals are reported as an AI channel rather than lumped into generic "referral"; Atlas sessions are recognized by their identifier and pulled out of direct; and AI visibility measures which answers the brand appears in. A diffuse "something with AI" thus becomes a named channel with its own conversion, its own dwell time and its own contribution to revenue.
Tuning analytics for agentic traffic in 2026
The changeover is less a project than a routine. Five steps bring reporting up to the state of the agentic web:
- Take stock: check how high your direct traffic is and how it has trended recently — a sudden rise with no campaign reason points to agentic traffic.
- Sort out bots: filter purely automated requests via network/ASN and automation markers out of the metrics, so the baseline is right.
- Name the agents: report agentic sessions (Comet, Atlas and others) as their own channel instead of scattering them across direct and referral.
- Measure first-party: move measurement into the cookieless first-party context so referrer, network and user agent arrive untampered.
- Watch the trend: track the AI channel over weeks in reports and status pages and treat an unexplained jump in direct traffic like an anomaly, not a coincidence.
Agentic traffic is not going away; it keeps growing. Building the separation of human, agent and bot into your reporting now means measuring in 2026 what actually happens — and not what a missing referrer pretends.
Start with a free trial and see how your direct traffic breaks down. Or book a personal demo to see bot detection, first-party measurement and the AI channel in action.
This article draws on data from: HUMAN Security ("State of Agentic Traffic", April 2026: agentic traffic growth around +6,900% since early 2025, agent browsers around 71% of agentic traffic, Comet around 48%, Atlas around 21%), Imperva ("Bad Bot Report 2025": automated traffic 51% of the web, bad bots 37%), Adobe Analytics (Quarterly AI Traffic Report 2026: AI-referred retail traffic +138% YoY in May 2026, +393% in the first quarter, +1,324% since October 2024, revenue per visit +37%, dwell time +48%, pages per visit +13%, engagement rate +12%, 39% use AI for shopping), MarTech (referrer behavior of Atlas and Comet), Seresa (around 71% of AI traffic as direct), SparkToro (30–60% of direct traffic misattributed, direct traffic around 37% of visits). Legal statements do not constitute legal advice.