Docs

AI Browsers and Dark Traffic: What Web Analytics Misses Now

AI browsers like Atlas and Comet present as standard Chrome to your analytics — track the signals that reveal agent traffic before it skews your funnel.

Acquisition

AI browsers such as ChatGPT Atlas and Perplexity Comet have moved from early-access curiosities to mainstream tools in 2026, and they change how your site gets visited in a way most web analytics setups aren’t built to see. These browsers present a standard Chrome user agent to every site they touch, so an agent reading your pricing page, filling a form, or comparing your product to a competitor’s shows up in your analytics indistinguishable from a human on desktop Chrome. Combine that with the fact that a large share of AI-driven visits arrive with no referrer at all — the assistant answered the user’s question and only opened your page as a side effect, or didn’t open it at all — and you get a growing pool of traffic that’s neither cleanly “direct” nor cleanly attributable to anything, just quietly inflating one bucket in your funnel report.

This matters most where teams make decisions off top-of-funnel numbers. A rising “direct traffic” line usually reads as a good sign — brand recognition, repeat visits, word of mouth. If a meaningful chunk of that rise is actually agent traffic from AI browsers reading your pages on a user’s behalf, you’re crediting the wrong channel and potentially misjudging which content is actually working. Conversion rate calculations get distorted the same way: if agents are landing on pages and leaving without the kind of interaction a human would have, blended conversion rate drops for reasons that have nothing to do with your funnel.

Data Points to Track

  • Session behavioural fingerprint — time-on-page, scroll depth, and interaction timing that differs sharply from typical human patterns, since user agent alone can’t separate AI browsers from real Chrome
  • Zero-referrer sessions as a distinct, tracked segment, rather than folded into “direct traffic” by default
  • Page-read-without-interaction sessions, where content loads and renders but no click, scroll, or form engagement follows within a reasonable window
  • Conversion rate calculated separately for flagged agent-pattern sessions versus everything else, so blended metrics don’t hide the split
  • Volume trend of unattributed/zero-referrer traffic over time, tracked weekly, to catch the shift as it grows rather than after it’s already large

Setup Steps

  1. Segment zero-referrer sessions into their own analytics bucket instead of letting them default into “direct,” and review that bucket’s size weekly.
  2. Add behavioural heuristics — sub-second-to-first-scroll timing, no-mouse-movement patterns, page loads with zero subsequent interaction — as flags on session records, not as a hard bot/human classifier, since false positives on real users are a real risk.
  3. Track this segment’s share of total sessions over the last 90 days, so you have a trend line before making any attribution or content decision based on it.
  4. Exclude or clearly separate flagged sessions from conversion-rate reporting used for marketing spend decisions, while keeping them visible for content-visibility purposes.
  5. Revisit the heuristics monthly, since AI browser behaviour and market share are both still changing quickly enough that a rule set from this quarter may be stale by next.

Actionable Insights

Once agent-pattern traffic is broken out, “direct traffic is up” stops being an ambiguous good sign and becomes two separate, useful numbers: real human direct visits, and a rough measure of how much AI-browser traffic is reading your site right now. The second number is worth watching in its own right — it’s an early proxy for how visible your content is to the assistants users are increasingly asking product questions through, even when that visibility never shows up as a click. Treat a rising agent-traffic share as a signal to check whether your key pages are structured clearly enough for an assistant to read and summarise correctly, not as noise to filter out and ignore.

Expert help

Need help tracking this in your app?

Our team sets up analytics pipelines for mobile and web teams every day. Talk to us and get your first events flowing in under an hour.

Talk to an expert