In one month, the agentic web became a two-company market.
HUMAN Security's State of Agentic Traffic report for October 2026, published on 9 November, finds that AI agents from Meta and OpenAI made about 90% of the agent requests reaching the sites it protects. Meta's Muse alone sent 59% (vendor-stated).
Total agent traffic in October was 2.6 times September's level and roughly eight times August's (vendor-stated). Media sites received 19.1% of it.
That last number is the one publishers should sit with. Agents are reading media sites at a rate that more than doubled in a month. Almost none of that reading is visible, priced or paid.
What is agentic traffic, and how is it different from crawler traffic?
Agentic traffic is the set of requests an AI agent makes while acting for a person in real time: loading pages, running searches, filling forms, occasionally checking out. HUMAN counts requests, not users or sessions. One request is one action, such as loading a page.
It is not crawler traffic. HUMAN says it does not infer agent activity from crawlers, which mostly serve model training, search indexing and chatbots.
It helps to keep three machine readers apart:
- Training crawlers collect pages in bulk to build models.
- Live Search Agents fetch a page the moment a user asks ChatGPT, Perplexity or Claude a question, and usually do not run JavaScript.
- Agentic browsers and built-in agents drive a browser to complete a task for the user. HUMAN's report measures this third group.
Each needs a different policy. A robots.txt rule written for a training crawler says nothing useful about Muse.
Who is sending the agent traffic?
HUMAN's October share of agent requests (vendor-stated):
- Meta Muse: 59%
- OpenAI ChatGPT Agent: 15.9%
- OpenAI Codex Browser: 8.8%
- OpenAI Codex Chrome extension: 5.2%
- Perplexity Comet: 3.6%
- Anthropic Claude Chrome extension: 3.5%
- Google Agent (Gemini): 0.8%
- xAI Grok Agent: 0.7%
Muse now sends 3.7 times as many requests as ChatGPT Agent, the next largest. Its share rose 19 percentage points in a month and its volume nearly quadrupled. OpenAI's three products together account for about 30%.
Why did Comet collapse while Muse surged?
Distribution. In August, Perplexity's Comet held 41% of agent traffic in HUMAN's data. By October it held 3.6%, and its request volume was down about 30% since August while the market grew roughly eightfold (vendor-stated).
HUMAN's reading is that consumers are moving away from standalone agentic browsers and towards agents built into platforms they already use. OpenAI shut its own standalone browser, ChatGPT Atlas, in August. Genspark's agentic browser dropped out of the top 10.
The longer arc is simple. Agent traffic was roughly flat through the summer of 2026, then took off in September when Muse arrived. The companies that already own the audience now own the agents that read the web for it.
What are agents doing on the sites they visit?
Mostly researching (all vendor-stated):
- 85.2% of agent requests went to product and search pages, up from 81% in September.
- Content engagement rose from 3.6% to 4.3%, now the second largest route.
- Checkout and payment fell to 1.9% of requests, but checkout volume grew 117% month on month.
By sector, ecommerce took 59.9% of agent requests, travel 19.3%, media 19.1% and SaaS 1.2%. Media's share slipped from 21.1% in September, yet its agent traffic still grew 2.3 times in a month.
One caveat matters. These figures describe traffic on applications HUMAN protects, which skew to commerce and travel. They are not a census of the web. HUMAN also sells agent detection, so it has a commercial interest in the problem looking large.
What does it mean for publishers?
The read is growing faster than any referral ever did. A 2.3x monthly jump in agent requests to media sites is reading at scale. An agent that finishes the task for its user does not send that user to the page.
Most of it is invisible. HUMAN says analytics tools on most sites cannot tell an AI agent from a human visitor (vendor-stated). OPA Europe and Rossel found 68% of media sites fail to detect basic bots (named-party figure). If you cannot see the read, you cannot price it.
Some of it may be miscounted as audience. Unlike Live Search Agents, agentic browsers drive a real browser, so page code and ad tags can run. An impression served to an agent is not human attention, and advertisers and verification vendors work to exclude non-human traffic. Revenue booked against agent sessions is fragile revenue.
Concentration cuts both ways. Two companies now account for about nine in ten agent requests. That makes the counterparties easy to name and policies easy to target per agent. It also means two firms set the terms of how media gets read.
Blocking is still a trap. Turning Muse away does not stop the user's question. It sends the agent to whichever source lets it in.
What does it mean for advertisers?
Discovery is moving into agents. With 85% of agent requests aimed at product and search pages, the shortlist is being built by software before a person sees it. A brand the agent cannot find or parse is missing at the moment of decision.
The dispatchers run ad businesses. Meta and OpenAI both sell advertising. The platforms sending the agents are positioned to sell influence over what those agents see and recommend.
Your numbers may be inflated. If agent visits are counted as human sessions, conversion rates and engagement metrics drift. Heading into Black Friday, that is worth checking before budgets move.
What should publishers do now?
- Separate the reads. Split agent requests from human sessions in server logs, not just in JavaScript analytics.
- Set policy per agent. Muse, ChatGPT Agent and Codex behave differently. Treat them as distinct counterparties, not one "AI" switch.
- Price the read. Decide what an agent request to your content is worth, and make sure it is not quietly billed to advertisers as a human impression.
the blankspace read The agentic web just consolidated into two dispatchers, and media is a fifth of where they go. That is not a traffic story; it is a supply story. Every agent request to a newsroom is a read that delivers value to Meta or OpenAI's user and nothing to the publisher. Referrals are not coming back to pay for it. The only durable answer is to monetise the read itself: see it, identify who sent it, and make it a priced, measurable unit rather than an invisible cost.
Key figures
| Figure | Value | Provenance |
|---|---|---|
| Meta and OpenAI share of agent requests, Oct 2026 | about 90% (89%) | vendor-stated (HUMAN) |
| Meta Muse share of agent requests | 59% | vendor-stated (HUMAN) |
| Muse requests vs ChatGPT Agent | 3.7x | vendor-stated (HUMAN) |
| ChatGPT Agent share | 15.9% | vendor-stated (HUMAN) |
| Agent traffic growth, Sept to Oct | 2.6x (157.5%) | vendor-stated (HUMAN) |
| Agent traffic growth, Aug to Oct | about 8x (697.5%) | vendor-stated (HUMAN) |
| Media share of agent requests | 19.1% (21.1% in Sept) | vendor-stated (HUMAN) |
| Media agent traffic growth, Sept to Oct | 2.3x | vendor-stated (HUMAN) |
| Requests to product and search pages | 85.2% | vendor-stated (HUMAN) |
| Checkout share / checkout volume growth | 1.9% / +117% MoM | vendor-stated (HUMAN) |
| Perplexity Comet share, Aug vs Oct | 41% vs 3.6% | vendor-stated (HUMAN) |
| Media sites failing to detect basic bots | 68% | named-party figure (OPA Europe / Rossel, Jan 2026) |
Primary source, opened and verified: HUMAN Security, "State of Agentic Traffic, October 2026: Meta and OpenAI now account for 90% of agentic traffic" (Aviad Kaiserman, 9 November 2026): https://www.humansecurity.com/learn/blog/october-2026-agentic-traffic/
