AI agents on the web now

For the first time in ten years, most of the traffic on the web is not people. Security firm Imperva measured automated software at 51% of all web traffic in 2024, and classed 37 points of it as outright malicious bots [1]. Most of that is dumb machinery — scrapers, spammers, ticket scalpers. But a growing share are true AI agents: software given a goal that works toward it on its own. This page is a tour of the agent families an ordinary person already meets online.

Support agents: the ones you've already talked to

If you've asked a company's help chat a question lately, there's a fair chance an agent answered. The clearest public example is the payments company Klarna. In its first month, February 2024, Klarna's AI assistant held 2.3 million conversations — two-thirds of all its customer-service chats — did work Klarna says once needed 700 full-time people, and cut the average time to solve a problem from 11 minutes to under two [2]. A whole market now rents this kind of agent to any company that wants one — that's agent as a service, covered on its own page.

Shopping agents

Amazon launched Rufus in 2024: a shopping assistant trained on Amazon's product catalog that answers questions like "what should I look for when buying headphones?" inside the shopping app [3]. It grew quietly huge — more than 300 million customers used it in 2025 — and in May 2026 Amazon folded it into a single "Alexa for Shopping" assistant across its apps, sites, and Echo speakers [4]. Agents also shop for you from the outside: OpenAI's Operator, launched as a research preview in January 2025, drives a real web browser to book tables, order groceries, and buy tickets, and has since been folded into ChatGPT's agent mode [5].

You no longer go looking for agents on the web. They're in the help chat, the shopping app, and the comment section.

Social media agents

In July 2024 Meta launched AI Studio, which let Instagram creators build an AI character — "an extension of themselves", in Meta's words — that answers their DMs and story replies, with every AI reply labeled so fans know [6]. (Meta stopped taking new characters in August 2026, though existing ones still run [6].) The honest caveat: labeled agents like these are a small island. Most automated accounts on social platforms are plain, undisclosed bots — part of that 37% of traffic Imperva classes as malicious [1]. Our rule of thumb — opinion, not science: if an account never sleeps and always sells, assume software.

Financial agents

The oldest mainstream agents on the web manage money. US regulators call them robo-advisers: "automated digital investment advisory programs" that ask you a questionnaire about goals and risk, then build and manage an investment portfolio for you, usually for lower fees than a human adviser [7]. They're regulated, unglamorous, and everywhere — which, in our opinion, is roughly where every other family on this page is heading.

If you want to sort these families more precisely, the types guide explains the difference between a scripted bot and a real agent — and the failure modes guide covers what happens when agents like these get things wrong.

Run a website? Agents are reading it too

Everything above is about agents you meet. The same shift runs the other way as well: agents now visit websites on someone's behalf, reading a page to answer a question or clicking through it to get a job done. Google's own guide for developers warns that sites designed for humans, "with complex hover-states, shifting layouts, and fluid motion," are "functionally broken for agents" [8].

That also changes the old advice to simply publish more. Our field guide Does publishing more get you found? walks through what Google actually documents: why one thorough page gets found by questions you never targeted, why a page for every keyword breaks its spam policy, and what agents need from a site that people barely notice.

AiAgentNook editorial · corrections are noted here, not silently edited.