Meta and OpenAI Agents Drive Nearly 90% of Agent Traffic

What happened

HUMAN Security's monthly State of Agentic Traffic report for October puts a number on who is sending AI agent traffic to websites. According to a write-up in Search Engine Journal, agents from Meta and OpenAI made nearly 90% of the agent requests across sites and apps HUMAN protects.

Meta's Muse alone accounted for 59%. It launched on Sept. 8, and its request volume nearly quadrupled in a month. HUMAN says it now sends 3.7 times as many requests as ChatGPT Agent, the next-largest agent.

OpenAI has three agents in the top 10 and about 30% of requests combined. ChatGPT Agent fell to 15.9% share, even though it sent 44% more requests than before. Muse simply grew faster.

Total agent requests in October were roughly 8 times August's level and about 2.6 times September's. This is HUMAN's customer footprint, not the whole internet, so treat it as a strong signal rather than a census.

Standalone agent browsers are losing ground

Perplexity's Comet accounted for 41% of agent requests in August. That fell to 9.4% in September and 3.6% in October. It was also the only top-10 agent whose request volume dropped, by roughly 30% since August.

HUMAN's explanation is that people pick agents from platforms they already know or products from the same company. Distribution beats novelty. Google's Gemini agent, launched Oct. 8, sits at No. 8 with 0.8%, and xAI's Grok Agent is at No. 9 with 0.7%.

Agents browse far more than they buy

In October, 85% of agent requests landed on product and search pages, up from 81% in September. Checkout and payment pages fell from 2.4% to 1.9% of requests.

That doesn't mean agents are irrelevant to revenue. HUMAN notes that checkout-bound requests are growing quickly, and that a request reaching a checkout page doesn't mean an order was placed. But right now, agents are mostly doing research on your catalogue and search results.

Ecommerce held steady at about 60% of agent requests. Travel came second, overtaking media for the first time in this series.

Why this matters for ecommerce teams

The uncomfortable part is identification. HUMAN says analytics tools on most sites can't tell AI agents from human visitors. Meta's own engineering post says Muse's browsing "will appear as your activity" to the sites it visits.

So your sessions, product views and on-site search data may already include agent behaviour that looks human. That can distort conversion rates, merchandising decisions and ad attribution.

There is also a technical wrinkle. Muse moved its Web Bot Auth signatures from meta.com to muse.ai between Nov. 7 and Nov. 9, per DataDome. As of Nov. 10, DataDome saw half of Muse's requests on its network arriving unsigned. A site that only trusts the meta.com directory would see Muse as unverified.

And because an agent like Muse has no name for robots.txt to match, the decision on access moves to your bot protection layer, not your robots file.

What to do before Black Friday

• Ask whoever runs bot protection how Muse and OpenAI's three agents are handled today. Don't assume the rules from a few months ago still apply.
• Check signature rules. If you verify Web Bot Auth, confirm the muse.ai key directory is included, not just meta.com.
• Decide your policy deliberately. Unsigned doesn't automatically mean malicious, and signed doesn't automatically mean welcome. Choose what you want agents to see, such as product, pricing and stock, and what you want protected.
• Review analytics with caution. Treat traffic spikes on product and search pages with some scepticism and look for patterns in server logs.
• Keep product data clean. Agents spend most of their time on product and search pages, so accurate prices, availability and structured data matter.
• Watch checkout separately. The share is small now, but it is the number that will show when agents start converting.

What to watch next

The key question is whether more of Muse's requests arrive signed. HUMAN says next month's report will cover Black Friday and Cyber Monday, which will be the first real look at agent behaviour in peak shopping season.