11 Nov 2026 · 4 min read

How Should Publishers Measure and Identify AI Agents Visiting Their Sites?

TL;DR: AI agents are a growing share of publisher traffic, but standard analytics was built to count humans, not agents retrieving a page once and leaving. A few concrete signals help estimate agent traffic until better tools catch up.

A new kind of visitor, and most analytics were not built for it

An AI agent, in this context, is software acting on behalf of a person, visiting a publisher's site to gather information, complete a task, or in some cases make a purchasing decision, without a human directly clicking through each page. This traffic has grown fast enough in 2026 that several publishers now treat it as a distinct audience segment rather than noise to filter out. The problem is that standard web analytics was built to count human visitors, and an agent visiting a page does not behave like one.

Why agent traffic doesn't show up the way human traffic does

Most analytics platforms infer a human visitor from a pattern of behaviour: mouse movement, scroll depth, time on page, a sequence of clicks. An agent retrieving a page to extract specific information does not generate that pattern. It may request a page once, pull the exact content it needs, and leave in a fraction of the time a human would spend reading the same content. Some agents identify themselves honestly through a user agent string. Many do not, either because the underlying model wasn't built with that convention in mind, or because the operator has no strong incentive to make agent traffic easy to separate from human traffic in log files that still get billed and reported the same way.

What publishers can actually check for

A few concrete signals help separate likely agent traffic from human traffic without needing a new analytics platform. Request patterns that don't match typical browsing, a single rapid page fetch with no subsequent navigation, are a reasonable first flag. User agent strings, where honestly declared, identify a known agent directly, though this undercounts the agents that don't declare themselves. Server logs, read independently of a JavaScript-based analytics tag, catch traffic that never executes the tracking script at all, which matters because some agents retrieve raw page content without rendering it the way a browser does. None of these signals alone is definitive. Together, they give a publisher a reasonable estimate rather than a precise count, which is itself a meaningful improvement over not measuring agent traffic at all.

Why counting matters beyond curiosity

A publisher that cannot distinguish agent traffic from human traffic cannot answer a basic commercial question: is an advertiser paying for impressions a human will see, or impressions an agent retrieved and a human never did. That distinction already matters for brand safety and viewability conversations, and it will matter more directly once more advertisers are specifically targeting agent-visible placements, a category some publishers have started testing deliberately rather than treating agent traffic purely as something to exclude.

The identification problem is also a trust problem

Once an agent is recognised as a distinct kind of visitor, a further question follows: which agent is it, and can its claimed identity be verified, or is it simply declaring itself. This is closely related to the work IAB Tech Lab is doing through its Agent Registry, where companies can list their agents, though the registry does not itself verify agent behaviour. For publishers specifically negotiating deals with buying agents rather than simply measuring their traffic, the same verification problem shows up again: a deal needs a record both sides can trust, not just an agent's self-reported identity. At Alkimi, the DealSheet addresses that narrower problem directly, once a deal is actually being negotiated rather than just measured, holding a timestamped record both the buying and selling agent can retrieve independently once terms are agreed.

Measuring agent traffic and verifying agent identity are related problems, not the same one, and a publisher working through the first is likely to run into the second within the next year rather than the next five.

Frequently asked questions

Why doesn't standard analytics count AI agent traffic accurately?

Analytics infers a human visitor from behaviour patterns like scroll depth and time on page. An agent retrieving a page once to extract information does not generate that pattern.

What signals can publishers check for agent traffic?

Unusual request patterns, honestly declared user agent strings, and server logs read independently of JavaScript-based tracking tags.

Does checking these signals verify an agent's identity?

No. Measuring traffic and verifying an agent's identity are related but separate problems. A declared user agent string is not the same as a verified one.

Why does this matter commercially, not just analytically?

A publisher that cannot distinguish agent from human traffic cannot answer whether an advertiser is paying for impressions a human will actually see.

All articles

Speak to the team

Request a demo

You don't need to plug in your agent, the demo walks you through the steps.