Agents built by Meta and OpenAI produced close to 90 percent of the agent requests that HUMAN Security recorded across its customers’ sites and apps in October. HUMAN’s State of Agentic Traffic report, published monthly, credits Meta’s Muse with 59 percent on its own. Search Engine Journal covered the findings on November 14.

The scale of growth matters as much as the split. October agent requests ran roughly eight times above August and about 2.6 times above September. HUMAN notes the report covers only traffic on its own protected properties, not the whole internet, and that each request is a single agent action such as loading a page or running a search.

Muse gained 19 percentage points of share from September, and its request count nearly quadrupled. HUMAN says Muse now sends 3.7 times the volume of its closest rival, ChatGPT Agent. Meta introduced Muse on September 8, so the product went from launch to dominant share in about seven weeks.

OpenAI holds about 30 percent across three agents in the top 10. ChatGPT Agent’s share fell from 28.5 percent to 15.9 percent even though its request count rose 44 percent, because Muse grew faster. The Codex Browser application rose from 7.7 to 8.8 percent, while the Codex extension for Chrome climbed from 3.3 to 5.2 percent.

Standalone agent browsers fared worst. Perplexity’s Comet held 41 percent of requests in August, then 9.4 percent in September, then 3.6 percent in October, with its volume down about 30 percent since August. The Claude Chrome extension slid from 7 percent to 3.5 percent of share while its volume grew 30 percent. Google’s Gemini agent, launched October 8 according to HUMAN, placed eighth at 0.8 percent. HUMAN attributes the pattern to users picking agents from platforms they already know.

Most of this traffic is browsing, not buying. Product and search pages drew 85 percent of requests, up from 81 percent, while checkout and payment pages slipped from 2.4 percent to 1.9 percent. HUMAN says the categories describe where a request landed, so a checkout hit does not prove an order. Ecommerce held steady near 60 percent of requests, and travel ranked second, passing media for the first time in the series.

The identification layer is where the operational risk sits. In Meta’s own engineering post, Muse’s browsing “will appear as your activity” to the sites it visits. DataDome reported on November 11 that Muse moved its Web Bot Auth signing keys from meta.com to muse.ai between November 7 and 9. As of November 10, DataDome counted half of Muse’s traffic across its network as unsigned, so a site checking only the old directory would see no verifiable identity.

Weigh the source carefully. Both HUMAN and DataDome sell products that identify bots and agents, and a finding that agents are hard to identify supports that business. HUMAN also says its method blends behavioral data, user-agent strings, and publisher-level integration, without breaking out how much of Muse’s count each method supplies. The figures describe HUMAN’s customer base, not the web.

The search angle is access control. Slobodan Manic argued in SEJ in October that robots.txt gives an agent like Muse no name to match, which pushes the allow-or-block decision into bot protection. HUMAN adds that analytics on most sites cannot separate agents from human visitors, so a spike in product-page sessions may be partly automated.

Ecommerce and travel teams should ask the owner of bot protection, ahead of Black Friday, how Muse and the three OpenAI agents are treated, and whether the muse.ai key directory is in the verification rules. HUMAN says its next report will cover Black Friday and Cyber Monday, which gives teams a benchmark to compare against their own logs.

Search Engine Journal, reporting HUMAN Security’s State of Agentic Traffic data, published November 14, 2026.