AI agent activity · Field data
What AI Agents Actually Do, Click by Click
Cequence followed 281,627 AI agent visits across five industries to see what each agent did, page by page. Most read content or compared products. Others reserved rental cars, checked out, and signed in to customer accounts, with two browser agents making nearly three-quarters of those sign-ins. And some were actually bot attacks masquerading as legitimate agents. Read on to see which agents handle which jobs, the paths they take, and how their activity changes from one industry to the next.
Source: Cequence field data, 281,627 AI agent visits across five industries, 15–21 Oct 2026.
Which agent did which job
Each band shows a share of visits, running from the agent on the left to the most purposeful thing it did during the visit on the right. Reading and looking things up accounted for 46% of visits, shopping and comparing for 39%, and signing in on the user's behalf for 2%. Two browser agents, Meta Muse and ChatGPT Agent, made 72% of those sign-ins.
Browser agents
Open the website and use it for the person: click, search, fill in forms, and check out.
Chatbot fetchers
Pull a page or two so a chatbot can answer the person's question.
Developer tools
Coding tools and other agents that fit neither group.
Browser agents and chatbot fetchers behave differently. Amazon's agent spent 92% of its visits shopping, while ChatGPT-User, Claude-User, and the other fetchers spent 60% to 68% of theirs reading.
Six common journeys, step by step
Each card shows a real path that repeated across many visits. Step names come from the page each request reached, and the cards leave out background requests for scripts, images, and analytics. The highlighted step names the job.
Amazon agent
Browser agent
Browses the catalog
·
Retail & E-commerce
Browse catalog
View product
Opens a category, then a product page. The agent's cart requests only checked cart status, so this analysis doesn't count them as adding to cart.
826 visits
took this path
Google-Agent
Browser agent
Goes to checkout
·
Retail & E-commerce
Add to cart
Check out
Read page
Adds an item, then runs the checkout flow, including the payment-token and checkout-process calls.
95 visits
took this path
Meta Muse
Browser agent
Works a reservation
·
Travel & Mobility
Book or reserve
Read page
Browse catalog
Read page
Book or reserve
Read page
Starts a reservation, reviews car classes and rental terms, then returns to the reservation.
407 visits
took this path
ChatGPT Agent
Browser agent
Finds a car
·
Travel & Mobility
Read page
Browse catalog
Book or reserve
Read page
Browses car classes, then starts a reservation.
63 visits
took this path
ChatGPT-User
Chatbot fetcher
Answers a question
·
Retail & E-commerce
Read page
Read content
Fetches a page and a related article so ChatGPT can answer the user's question.
414 visits
took this path
Claude-User
Chatbot fetcher
Checks the site rules first
·
Financial Services
Read site rules
Read page
Reads the site's rules file before it fetches the page the user asked about.
96 visits
took this path
414K
Attackers pretended to be Google-Agent to run credential stuffing
At one online marketplace, 414,210 visits identified themselves as Google-Agent. Every one went straight to the sign-in page, went nowhere else, and returned errors. That pattern is credential stuffing from attackers using Google-Agent's name, so this report excludes those visits from Google-Agent and from every other figure on this page. Had they been counted, sign-in would have made up 60% of all agent activity.
What agents do in each industry
The inner ring shows the job, and the outer ring breaks each job into tasks, shaded from the job's color. Tasks under 3% merge into an unlabeled slice. Travel & Mobility stands apart: booking makes up 45% of agent visits there, while agents on Financial Services sites spend 86% of their visits reading.
The task-level split inside reading comes from one day of page data (21 Oct) and is an estimate. The analysis leaves out background calls made by each site's own app.
About this data
The figures in this report come from AI agent traffic that the Cequence Platform observed across customer web applications and APIs in five industries between October 15 and October 21, 2026. Cequence identifies AI agents in three ways. Some announce themselves in the user agent as visitors acting for a person, such as ChatGPT-User and Claude-User. Some declare an agent identity in a request header, following the Signature-Agent convention, as ChatGPT Agent and Google-Agent do. Others don’t announce themselves at all, and Cequence recognizes them by the browser, operating system, and network pattern they present, as with Meta Muse.
Cequence then grouped each agent’s requests into visits and read the pages each visit requested, in order, to decide what job the agent was doing. The analysis left out training crawlers and search indexers, background requests for scripts, images, and analytics, and calls that each site’s own app makes. It also left out traffic that claimed to come from an AI agent but behaved like an attack, such as visits that only reached sign-in pages and failed. Before analysis, Cequence removed anything that could identify a customer, an end user, or an individual request, including account details, credentials, IP addresses, and the content of forms. The report shows only combined counts and percentages by agent, job, and industry. No figure on this page can be traced to a specific organization or person.