- HUMAN Security measured a 209.7% rise in agentic traffic in September over August, led by Meta's Muse and OpenAI's ChatGPT Agent.
- Most agent activity is research, not purchase, and large platforms are choosing opposite responses: Amazon has blocked Muse's agents, while Shopify gives agents their own checkout lanes.
- Leaders can ask whether their analytics separate agents from people, and who decides which agents may read, log in or buy.
For years, a website had two kinds of visitor: people and bots. Each had its own rules. Personal AI agents fit neither box. They act for one person, yet they move at machine speed. The lesson here is that access looks less like a switch and more like a policy, shaped agent by agent and action by action.
What HUMAN measured
HUMAN Security publishes a monthly benchmark of agentic traffic seen across its Sightline Cyberfraud Defense and AgenticTrust products. It reports that agentic traffic rose 209.7% in September over August. It calls this the largest single-month increase it has recorded this year.
For scale, March was the strongest of the previous seven months, and it added 43.4%.
Two launches drove the jump. Meta announced Muse on September 8. By month's end it was the biggest agent in HUMAN's data, with 40.3% of the month's agentic traffic. OpenAI's ChatGPT Agent reached 4.5 times its August traffic, and its share rose from 19.8% to 28.6%. Comet Browser, previously first, fell from a 40.7% share to 9.4%.
MIT Technology Review separately cites a figure of 5 million downloads for Muse in its first month. It also says agentic web traffic tripled from August to September.
One caveat matters. HUMAN's figures come from one vendor's platform and detection methods. HUMAN also sells agent-control tools. Read them as a benchmark, not a census.
How a site tells an agent from a person
HUMAN says it identifies agents through three signals: behavior, the name a visitor announces about itself (the user-agent), and direct integration with publishers where available. It also says most analytics tools cannot tell AI agents from human visitors. That is a vendor's claim, but it is easy to test on your own reports.
The default setting is part of the story. HUMAN's Sightline blocks Muse traffic by default. Each customer must choose to allow it. HUMAN says many are still blocking it, and blocked requests reached a new high in September. If you use a bot-defense product, check whether it blocks Muse by default and who approved that.
Agents browse far more than they buy
HUMAN says product and search routes made up 81% of agent activity in September. Checkout and payment made up 2.4%. HUMAN concludes that agents are used mainly for research, not for completing transactions.
Ecommerce took 60.4% of agentic traffic, up from 46% in August. Media fell from 36.2% to 21.1%. HUMAN notes that sector labels reflect the page visited, not the agent's intent.
Muse users look cautious. Account routes were 1.7% of Muse traffic against 4% across all agents. Login routes were 1.1% against 3.1%. HUMAN suggests limited early trust as the likely reason. That is its interpretation, not a measured cause.
Large platforms are choosing opposite answers
MIT Technology Review reports that Amazon has blocked Muse's agents. Amazon is also fighting the AI startup Perplexity in court to block its agents. A judge recently ruled for Perplexity, and Amazon has appealed. Amazon argues that agents make it harder to tell ads seen by humans from ads served to bots.
Shopify has gone the other way. MIT Technology Review reports it gives agents their own checkout lanes that bypass pages built for humans. The same report says TechCrunch found agents turned away from sites including Pizza Hut and Yelp.
Both choices follow business logic. Amazon's argument is about how advertising is measured. Shopify's is about who is allowed to buy. The Wharton professor Ethan Mollick argues that agents sent to seek refunds or negotiate prices could overwhelm customer service. That is his view, not a finding.
Questions to put to your team
Can our analytics separate agents from people today, and who checked? Which agents are blocked by default, and who approved that? For each agent, which actions are allowed: reading, adding to cart, logging in, paying?
If some ad viewers or shoppers are agents, how do we count impressions and conversions? Does customer support have a path for requests made by an agent on a customer's behalf?
A gate built for two kinds of visitor may now be setting policy for a third. The question for leaders is who owns that choice.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: HUMAN Security.
