AI Browsers Just Arrived in Your Analytics: What Atlas Means for Ads

In the space of a month, three of the biggest names in AI shipped browsers that can browse, click and fill in forms for the person using them. AI browser traffic is now reaching marketing sites, and it arrives looking like an ordinary visitor on Chrome. This article covers what launched, how these sessions show up in Google Analytics 4 (GA4) and on paid campaigns, which patterns to watch for, and what to do this quarter so you have a clean baseline before the volume grows.

What launched: Atlas, Comet and Copilot Mode

The category went from niche to mainstream in October 2025:

• Perplexity Comet, October 2. Perplexity made its Comet browser free for everyone worldwide, after a limited release that began in July. Its Comet Assistant browses alongside the user, and Perplexity also announced Background Assistants that work on tasks asynchronously.
• OpenAI ChatGPT Atlas, October 21. OpenAI introduced ChatGPT Atlas on macOS, with Windows, iOS and Android to follow. Agent mode, in preview for Plus, Pro and Business users, lets ChatGPT research, compare and complete tasks in the browser while you watch.
• Microsoft Edge Copilot Mode, October 23. Two days later, Microsoft expanded Copilot Mode in Edge with Actions, which can take steps like booking a hotel or filling out forms on the user's behalf.

The common thread matters more than any single product. These agents do not run on a server farm with a crawler name in the header. They run inside a full browser on the user's own machine, and they load pages, run your JavaScript and click on things the way a person would.

A web that was already mostly bots

These agents arrive on a web that was already mostly automated. The 2025 Imperva Bad Bot Report found that automated traffic reached 51% of all web traffic in 2024, surpassing human activity for the first time in a decade, and that malicious bots made up 37% of all traffic. Imperva also warned that accessible AI tools have lowered the barrier to building bots at scale.

The fight over agents has already started

The line between a helpful agent and an unwanted one is also contested. In early November, Amazon sued Perplexity, alleging that Comet's shopping agent was disguising automated activity as a human Chrome user. Perplexity responded that Comet only takes actions when a user asks it to. Whatever the court decides, the dispute shows that even the largest sites are struggling to tell agents from people, and that the question of whether to allow them is now a business decision.

Why AI browser traffic looks human to your analytics

Traditional bots are relatively easy to spot because they announce themselves or behave like scripts. AI browsers do neither.

The user agent is plain Chrome

On launch day, developer Simon Willison checked the Atlas user-agent string and found it identical to the latest Google Chrome on macOS. Atlas is built on Chromium, so to a web server it reads as Chrome. A marketer who tested Atlas against GA4 reported the same thing in the reports: in KP Playbook's walkthrough of Atlas in Google Analytics, Atlas shows up as "Chrome" in the browser dimension, not as its own browser.

Your analytics tag runs normally

Because the agent drives a real browser, your GA4 tag, ad pixels and consent banner all load. The session is recorded, events fire and, if the visit started from an ad, the click was counted too. Nothing in the pipeline has a reason to question it.

Known-bot filtering was not built for this

GA4 removes known bots automatically, using Google research and the Interactive Advertising Bureau (IAB) International Spiders and Bots List. According to Google's documentation on known bot-traffic exclusion, you cannot turn this off and you cannot see how much was excluded. A list of declared crawlers does little against a visitor that declares itself as Chrome.

Some agent operators are working on identification. In August, Cloudflare launched a "signed agents" program in which agents cryptographically sign their HTTP requests using Web Bot Auth; its first cohort included OpenAI's cloud-hosted ChatGPT agent, Browserbase and Anchor Browser. That helps site owners who use Cloudflare and can read those signatures. It does not change what GA4 or an ad platform shows you, and it depends on each agent opting in.

What AI browser traffic does to paid campaigns

For paid media, the problem is less about volume today and more about what these sessions do to the numbers you optimize against.

• Clicks without intent to convert. When a user asks an agent to research or compare vendors, the agent may open several of your pages, including ones reached through paid links. You pay for the visit, but no human saw your landing page.
• Distorted engagement and conversion rates. Agent sessions sit in the same Chrome bucket as your real visitors, so they drag down engagement and conversion rates in ways that look like a landing-page or audience problem.
• Bad signals for automated bidding. Smart bidding learns from the sessions and conversions you send back. If some of those sessions are agents, the model learns from behavior that no buyer produced.
• Attribution gaps. KP Playbook found that Atlas search traffic was attributed as chatgpt.com / referral or sometimes chatgpt.com / (not set), and that GA4 cookies did not carry over when Chrome data was imported into Atlas, so returning visitors appeared as new users. Both effects blur the channel and user counts that marketing attribution depends on.

Google defines invalid traffic (IVT) as activity that does not represent genuine interest in your business. An agent acting for a real person sits in a gray area: there is a human behind it, but the click on your ad is not that human looking at your ad. Ad platforms were not designed for that distinction, and advertisers should not assume it is being filtered for them.

Tell-tale patterns of AI agent traffic

There is no single field that says "agent" in GA4. You have to look at behavior. Dataslayer's launch guide for marketers on Atlas expects agent-mode sessions to show rapid page progression (five or more pages in under 10 seconds) and high click rates with near-zero engagement time. Treat those as hypotheses to test against your own data, and look for them together:

• Several pages viewed in a few seconds, often in a logical research order such as pricing, features and comparison pages.
• Very short engagement time per page, but deep scrolling or clicks on every link in a section.
• Form fields completed at machine speed, or partial form interactions that stop at a login or payment step.
• Sessions from "Chrome on macOS" that spike without a matching change in campaigns, budget or seasonality.
• Paid sessions with normal click-through but conversion rates well below the rest of the campaign.

None of these alone proves an agent. Fast readers exist, and so do old-fashioned bots that also claim to be Chrome. The value is in the combination, and in comparing against a period before agents were common.

What to do now: an AI agent traffic checklist

The best time to measure the change is before it becomes large. Work through these steps this quarter:

1. Record a baseline. Export the last 90 days of sessions, engagement time, pages per session and conversion rate by browser, operating system and channel. Note the date each AI browser launched so you can compare before and after.
2. Build a "suspected agent" segment. In GA4 explorations, filter for sessions with high page counts and very low engagement time. Watch its share over time, especially in paid channels.
3. Check your server or content delivery network (CDN) logs. Analytics only sees what your tag records. Logs show request timing, IP ranges and, where supported, signed-agent headers that never reach GA4.
4. Compare ad-platform clicks with on-site sessions. Reconcile clicks with landing-page sessions per campaign. A growing gap, or sessions that never engage, deserves a closer look. Our guide to spotting click fraud in PPC campaigns covers the reconciliation steps.
5. Protect conversion signals. Before sending conversions back to Google or Meta, make sure form submissions and leads come from verified human sessions, so bidding models do not learn from automation.
6. Decide your policy. Some agent visits may be welcome, for example a buyer's assistant comparing your pricing. Others, like automated form submissions, are not. Decide which you want to allow, measure and block, then apply it consistently.
7. Revisit monthly. Atlas is only on macOS today, with other platforms announced. Expect the mix to change quickly as it expands.

How Ðeny helps you see and control automated visits

Ðeny classifies traffic in real time, at the moment a visitor lands on your marketing site, rather than after the numbers have reached your reports. Each visit is assessed on behavior and network signals, not just the user agent, so traffic that claims to be Chrome is judged by what it actually does. Ðeny Bot Shield can block unwanted automated visits before they submit forms or trigger conversion tags, and persistent UTM attribution with a per-channel breakdown shows which campaigns send the cleanest traffic. You can see how the pieces fit together in the Ðeny features overview.

Conclusion

AI browsers turn the agent from a server-side crawler into a visitor that looks like one of your customers. For paid media teams, that means sessions and clicks that GA4 and ad platforms count as human, but that no human saw. The practical response is not panic but measurement: set a baseline now, watch the behavioral patterns, and make sure the signals you feed back to your bidding models come from real people. If you want to see how much of your paid traffic is human today, request a Ðeny demo.