What AI Agents Actually Do on Your Website

Personal AI agents have moved from demos to production traffic. People now ask an assistant to compare car insurance, fill in a quote form or check a policy, and the agent opens a browser and does it for them. For the businesses on the other side of the screen, that raises practical questions. How many of these visitors are there? How do they behave? Do they get stuck in the same places humans do? And can you even tell them apart?

Between 22 June and 29 September 2026, InfiniteWatch analysed 16 million web sessions across the websites of insurers, brokers and fintech companies that use our platform. This post shares what we found about how AI agents actually navigate, click and convert on real customer journeys.

Agent traffic tripled in three months, from 4% to 12%

In the first week of July, about 4% of classifiable desktop sessions came from AI agents. By the last week of September it was 12%, roughly 1 in 8 visits. The growth came in two steps, each a few days after a major consumer agent launch.

Instinct: from zero to a quarter of agent traffic in five weeks. The invite-only agent went viral in private beta in the third week of August. We first saw it on client websites on 25 August, a few days later, and its share climbed every week after that to 3.4% of all sessions.

Muse: the steepest jump. Meta's agent had been visible at low levels since July, most likely early testers. After its public launch on 8 September, its share jumped by half within a week and nearly doubled within two, reaching 5.7% of all sessions. Launching to the whole US market, inside apps people already use, made its adoption far more viral.

Everything else grew slowly. Headless automation and unidentified agents crept up from 2.4% to 2.9% over the period.

Most sessions never reach a verdict, and that is by design. Very short visits and mobile sessions are excluded for now, because the signals we rely on (cursor paths, click timing) need a desktop session with enough interaction to read.

Meet the agents: four families, four fingerprints

By late September, Muse alone accounted for almost half of all agent traffic, and Instinct for more than a quarter. We group agents into families by how they behave, because each automation stack leaves its own recognisable fingerprint.

Muse is the one to watch. It works in two modes. In the first, it teleports a cursor straight onto a button and clicks, leaving a single mouse sample per click where a human leaves a trail. In the second, it skips the cursor entirely and triggers buttons from code: every click registers at screen position (0,0), and several can land within milliseconds of each other on different buttons.

Headless automation is a broad bucket. It is a script driving an invisible browser, which could be an AI agent, but could also be a price scraper, a QA test or an uptime monitor. We identify it mostly from its technical signature rather than its behaviour.

Unknown agents are the most interesting group for the future. They behave like machines, yet match none of the fingerprints we have catalogued. Many are likely new agents, or new versions of known ones, and we review them to build the next generation of fingerprints.

Instinct is the newest arrival and, after Muse, the fastest riser. Its pattern is unmistakable: it treats every page like a checklist, centring each target on screen before clicking.

What an agent session looks like, click by click

An agent answers a quote form in the time it takes a human to read the first question. To see this up close, we ran agents and people through the same car insurance quote flow on our own demo website and recorded every event.

Here is an excerpt from an Instinct session, as it starts the quote:

1. 1.3.54 s: scrolls the page so the country buttons sit in the middle of the screen.
2. 2.3.65 s: presses "United States". The button is held for 7 milliseconds.
3. 3.3.90 s: scrolls again, placing the age buttons at the exact same vertical position.
4. 4.4.00 s: presses "26 to 35", and the postcode field is filled 4 milliseconds later.

Four answers in under half a second, with no cursor movement at all, and every click landing on the same horizontal line of the screen.

A person on the same form behaves very differently. Our human testers took a median of 40 seconds and 15 clicks to complete the flow. Their cursors produced around 75 movement samples per minute, drifted toward each button before clicking, and kept fidgeting during pauses while they read. Humans who clicked the same button twice almost never hit the same pixel again. Agents almost always did.

Agents are fast, focused and just as likely to reach the finish line

Agents spend less than half as long on a website as humans, yet reach conversion pages (quote results, checkout, application steps) at the same rate: 45% versus 46%.

They are efficient, not casual browsers. Agents view fewer pages and make fewer clicks. They skip the reading, the comparing and the second thoughts. An agent does not wander through your blog or reread the coverage details; it goes where the task sends it and stops.

They are high intent traffic. This is the finding that matters most commercially. Nearly half of agent sessions reach a conversion page, the same as people. These are not scrapers passing through. They are delegated buyers doing real work on behalf of real customers.

They hit errors slightly more often. 53% of agent sessions run into at least one error, compared with 49% of human sessions. Forms, validation messages and widgets were designed for eyes and fingers, and agents sometimes trip over them.

But they never look frustrated. Agents almost never rage click (0.2% of sessions versus 1.4% for humans). A person who gets stuck hammers the button; an agent quietly retries, gives up or takes another path. That means the frustration signals most analytics teams rely on are blind to agent friction. An agent can fail on your quote form and leave no visible trace of annoyance behind.

How we tell agents from people

In our controlled tests, a behavioural score flagged every agent that completed at least five clicks, and flagged none of the humans. It typically reached its verdict by the fifth click, about nine seconds into the session.

Many agents announce themselves through their technical signature, but the ones that matter most increasingly do not. So instead of relying on what a browser says it is, we look at how it moves. The strongest signals are things a human hand simply does not do:

• No approach path. A person moves the cursor toward a button before clicking. Agents click with no movement in the second before, or teleport the cursor straight onto the target.
• Pixel perfect repeats. Agents click the same spot of the same control every time. Humans almost never land in exactly the same place twice.
• Screen centre lock. Some agents scroll every target to the vertical middle of the screen and click on that exact line.
• Machine timing. Button presses shorter than 20 milliseconds, or a perfectly constant 80 milliseconds, click after click.
• Impossible speed. Two clicks on different buttons less than 50 milliseconds apart, or a text field focused and a different button pressed within the same instant.
• Silent pauses. When people stop to think, their cursor still drifts. When agents pause, nothing moves at all.

No single signal decides the verdict. Each one adds points to a combined score, and a session is labelled an agent only when the evidence crosses a threshold. This keeps false positives down: keyboard users, accessibility tools and fast typists can trip one signal, but they rarely trip several at once.

We also keep ourselves honest. When a new agent family appears, we first measure how well our existing rules catch it, and any new rule it inspires must be validated on fresh sessions it has never seen before we trust it. The same approach let us spot the "unknown agent" group: sessions that fail every human test but match no fingerprint we already know.

What this means for insurers and regulated businesses

If 1 in 8 of your desktop visitors is already an agent, a share that tripled in a single quarter. Since agents reach conversion pages as often as people, they are already part of your funnel. Treating them as bots to block would mean turning away customers who simply chose to delegate. Here is what we recommend.

1. 1.Measure agent traffic separately. Blended metrics hide it. Agents shorten average session times, lower pages per visit and dilute engagement numbers, which can make a healthy funnel look worse (or a broken one look fine).
2. 2.Stop relying on frustration signals alone. Rage clicks and dead clicks were built for humans. Agents fail silently, so you need agent specific friction detection: where they stall, retry, abandon or hit an error.
3. 3.Make your key flows agent ready. Clear labels, predictable buttons, forms that validate without visual only cues and error messages written in plain text all help agents finish the job. The same changes usually help accessibility too.
4. 4.Decide your policy, family by family. A personal assistant getting a quote for its owner is not the same as a scraper harvesting your prices. Knowing which family is on your site lets you welcome one and throttle the other.
5. 5.Think about compliance. In insurance and financial services, who submitted an application and what they saw matters. When an agent completes a form on someone's behalf, your records, consents and disclosures need to reflect that.

The shift is already underway. The businesses that understand how agents behave on their websites today will be the ones agents can actually buy from tomorrow.

About the data

The weekly trend (4% to 12%) and the late September family shares are measured on classified desktop sessions and shown against the public launch dates of Instinct (viral private beta, third week of August) and Muse (8 September). Each launch shows up in the data with a few days of adoption lag.

• Period and scope: 16 million web sessions on client websites in insurance, financial services and adjacent regulated sectors, 22 June to 29 September 2026. Test organisations are excluded.
• What we classify: desktop sessions with enough interaction to judge. Mobile sessions and very short sessions are not yet classified, so the true agent share across all traffic is unknown.
• Conservative by design: 1,359 borderline sessions were labelled uncertain rather than forced into either group.
• Error and rage click metrics use standard insights designed for human sessions. Agent specific problem detection is still in development.
• Headless automation may include scripts that are not AI agents, such as monitors or tests.
• Muse is labelled by our detection as associated with Meta, based mainly on technical signatures combined with behaviour.