HUMAN Security published its monthly “State of Agentic Traffic” report this month, covering October 2026. Meta and OpenAI agents made nearly 90% of the agent requests on the sites and apps HUMAN protects. Meta’s Muse, which launched on 8 September, made 59% of them on its own. Muse sends 3.7 times as many requests as ChatGPT Agent.

The report also shows where those requests went. 85% landed on product and search pages, up from 81% in September. Checkout and payment pages got 1.9%, down from 2.4%.

The claim: agents are about to do your shopping

The excitement runs like this. Shoppers will hand a task to an agent, the agent will buy, and your checkout had better be ready. Agent request volume does look big. Total requests in October were about 8 times August’s.

So is there a real need to rebuild checkout for agents? I ask that first with any new technology. The report says no, at least for now.

What the numbers say and what they don’t

Be careful with the 1.9%. HUMAN says the page categories show where a request ended up, not what the agent did. A checkout-page request doesn’t mean an order was placed. The report also covers only sites and apps HUMAN protects, not all internet traffic, so it won’t match your site exactly.

Even with those limits, the direction is clear. Agents spend nearly all their requests loading product pages and running searches. Very little of it reaches the pages where money changes hands. They are reading, comparing and narrowing down. The purchase, where it happens at all, comes after.

The other thing that changed is who is doing the reading. In one month the mix shifted a lot. Muse’s share rose 19 percentage points. ChatGPT Agent’s share fell from 28.5% to 15.9%. Perplexity’s Comet went from 41% in August to 3.6% in October. I would not build a plan around any one agent’s name, because the list moves too fast.

What this means for a small business

Take a shop that sells skincare, or a clinic that lists its treatments, or a service business with a page for each service. A shopper asks an agent for a recommendation. The agent visits your pages, reads them, and decides whether you make the shortlist. Your checkout never enters the picture if your product page gave it nothing to work with.

My position on agentic commerce is that it is too early to call who wins. The big platforms may block agents or open up to them, and they can afford to wait. A smaller business can’t wait, so the decision is already made for it. Get data-centric. Your website is one surface of many, and wherever your information shows up, it needs more precision and more detail than before.

An agent recommends you when your page gives it attributes that are materially different from your competitors’. “High quality” and “trusted by customers” give it nothing to compare. Sizes, ingredients, prices, lead times, who the product suits and who it doesn’t suit give it something to quote. The same goes for the odd questions. Everyone asks their agent the same thing a little differently, so your pages need to cover a wider spread of questions with varied, detailed answers.

Finding out who is reading

The report notes that analytics tools on most sites can’t tell AI agents from human visitors. So start with your server logs and any bot-protection reports you already have. Look for the agents named in the report and see which pages they request.

Expect it to be messy. Meta says Muse browsing “will appear as your activity” to sites it visits. Muse also moved its Web Bot Auth signatures from meta.com to muse.ai between 7 and 9 November, and DataDome reported that about half of Muse’s requests on its network were unsigned as of 10 November. Some of this traffic will look like ordinary visitors. Both HUMAN and DataDome sell agent identification services, so keep that in mind when reading their advice.

Where this shows up in my own work

At AeroChat, we publish many pages that each target one integration with a specific piece of software. Each page catches one narrow question. I would rather have a hundred precise pages than a few broad ones, because agents ask narrow questions.

What to do this week

  • Pull your server logs for October and check which agents visited, even if the picture is incomplete.
  • List the product, service and search pages they read most.
  • Rewrite those pages with specific facts: measurements, prices, timings, who it is for and who it is not for.
  • Add pages for the narrow questions customers ask you, one question per page.
  • Leave checkout alone unless your own data shows agents completing orders on it.
  • Check again after Black Friday. HUMAN’s next report will cover agent behaviour over Black Friday and Cyber Monday.

Source: Search Engine Journal: Meta’s Muse Now Sends Most AI Agent Requests One Security Firm Sees