Bots inflate bounce rates, AI agents slip past JavaScript tracking, and agentic browsers mix human and machine activity in one session. Snowplow's Nick Stanchenko explained how to detect each, combine server logs with analytics, and adapt websites to who is browsing.
Key takeaways
- Don't just block agents. AI agents bring users and revenue, so detect, analyze and adapt instead.
- Bots skew key metrics. They inflate page views and bounce rates and drag down click-through and conversion rates.
- Combine three signals. Who a visitor claims to be, where it comes from and how it behaves together separate bots from people.
- Most AI traffic is invisible. Many AI agents don't run JavaScript, so server logs are needed to see them.
- Sessions are now mixed. Agentic browsers put a human and an agent in the same session, which breaks metrics like engaged time.
More and more of the traffic hitting company websites isn't human, and much of it is distorting the numbers analytics teams report. In the Data Science Connect webinar "How AI Agents and Bots Are Breaking Your Web Analytics — And What to Do About It," Nick Stanchenko, Senior Product Manager at Snowplow, laid out three ways AI agents and bots break web analytics and what teams can do about each.
The webinar was sponsored by Snowplow, which describes itself as a customer data infrastructure platform. Brian Mink, co-founder and president of Data Science Connect, introduced the session. According to the session description, which cites HUMAN Security, agentic browsing traffic grew 6,900% between 2024 and 2025, driven by tools such as Perplexity Comet and ChatGPT Atlas.
Why can't companies just block AI agents and bots?
Stanchenko opened with figures he attributed to Cloudflare: 52% of HTML requests came from agents and bots, about 48% from non-AI agents and 4% from AI agents. People are also changing how they browse, he said, turning to chatbots and AI platforms instead of classic search.
The traditional response was to block bots with CAPTCHAs and firewalls. That no longer works, he said, "because agents actually bring you users and they bring you revenue." Snowplow proposes a three-stage framework instead: detect whether a visitor is human or agent, model and analyze their behavior separately, and adapt the website and content to whoever is browsing.
He then broke the problem into three parts. Older scraper bots run JavaScript tracking code and show up in analytics. Many newer AI agents don't run JavaScript and are invisible. And agentic browsers mix human and agent activity in the same session.
How do bots skew web analytics metrics?
Bots that run tracking code tangle themselves into human data, Stanchenko said. They tend to bounce after a single visit and rarely click or convert. The result is page views and bounce rates that are too high, and click-through and conversion rates that are too low. The fix, he said, is not to block them from the site but to remove them from the data being analyzed.
He described three signals, ordered from easiest to hardest for a bot to fake:
- Who they appear to be. A declared crawler is probably what it says it is, and signs of automation or unrealistic device sizes point to a bot. But a visitor claiming to be ordinary Chrome may still be a scraper.
- Where they come from. Snowplow recommends checking the autonomous system number behind an IP address. Traffic from a regular internet provider is probably human, while a "regular browser" coming from a cloud provider is probably a bot, though VPNs complicate this.
- How they behave. Mouse movement, scrolling and tapping suggest a person. Rapid page views at regular intervals suggest automation. Bots could mimic humans, he said, but slowing down defeats the purpose of the automation.
He cautioned that downstream filtering won't catch every bot, so upstream measures "are not going anywhere," but analytics still need cleaning by combining all these signals.
How can teams measure invisible AI traffic?
Stanchenko walked through how one AI platform touches a site. A training crawler adds a page to its data set. Months later, when a user asks a related question, a separate agent fetches the current page. Only then might the user click through, carrying a referral tag. Typical analytics show only that last step, "the tip of the iceberg."
Without the rest, he said, teams can't answer which AI platforms are active on their site, which send the most users, or which pages work well for AI. Clean analytics, he said, is "the first step to any sort of optimization," including AI engine optimization.
Server logs capture every visit because they don't depend on JavaScript. Stanchenko argued against sending those logs to a separate AI analytics tool, since questions such as how active AI platforms are on the most engaging pages need both data sets. He recommended one platform as a single source of truth for client-side events and server-side logs.
Asked whether AI referral attribution is consistent across platforms, he said, "I wish the answer was a resounding yes, but it isn't." Some platforms set a UTM source, some set a referrer header, some set both and some set neither. He expects attribution to improve as platforms seek content deals with media companies and need to show the audience they bring.
What do agentic browsers mean for analytics?
Analytics has long assumed a session is either human or agent, Stanchenko said. Agentic browsers break that assumption. He showed Perplexity's Comet taking over a browser to fill in a flight booking. A person might start on a carefully designed booking flow, then hand the task to the browser, which skips the pop-ups and promotions and goes straight to business. Engaged time looks normal for the human part of the session and near zero for the agent part.
Detection here relies on behavior: whether there is human activity and how the visitor navigates between pages. In a recorded demo on a mock website, Snowplow's real-time profiles flagged the agentic browser as not human after a few pages without human-like interaction. The site then showed the agent only a summary of each article, prompting the person to step in for the full content, and offered its own expert agent once a human returned. "We set our own terms for how the agent intermediates between our website and the user," he said in the demo.
He summarized three steps: untangle users from agents and bots, surface hidden AI traffic, and get ready for agentic sessions. With Chrome adding agentic browsing features, he said, browsing "will probably just get more agentic from this point onwards."
More sessions on data and AI
Data Science Connect has 24 free live webinars on its 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule for upcoming sessions. Registration for 2027 events is coming soon.
Frequently asked questions
How do AI bots affect web analytics?
Bots that run tracking code inflate page views and bounce rates and lower click-through and conversion rates, while many AI agents don't run JavaScript and never appear in analytics.
How can you detect bot traffic on a website?
Stanchenko recommended combining the browser's declared identity, the network the visitor comes from (such as a cloud provider's autonomous system number) and behavior such as mouse movement and scrolling.
How do you track AI agent traffic that analytics misses?
Server logs record every visit, including AI crawlers and agents; Stanchenko recommended loading them alongside client-side events in a single platform.
Is AI referral attribution consistent across AI platforms?
No. Some platforms set a UTM source, some a referrer header, some both and some neither, according to Stanchenko.
Watch: How AI Agents and Bots Are Breaking Your Web Analytics — And What to Do About It on YouTube
Event: Webinars · More coverage: Webinars insights
