Agent traffic needs its own line in your analytics

Contentsquare's October 13 relaunch makes a claim worth acting on whether or not you buy the product: agent traffic now needs its own line in your analytics stack. The new platform separates LLM and agent sessions from human sessions, and its attribution model credits AI-referred touchpoints instead of only the last click. The vendor published no figures on how much of any site's traffic that is, so the useful move is to measure your own.

Most marketing reports still treat every session as a person. If AI assistants and autonomous agents are visiting your site, researching on a buyer's behalf and referring traffic, those sessions are blended into averages that describe nobody. Segment first, then decide what to budget around.

What Contentsquare shipped

According to the release, the human, LLM and agent analytics "separates and analyzes LLM and agent traffic alongside human sessions", with the stated aim of optimizing for LLM referrals and agentic commerce. It adds what the company calls automated WebMCP monitoring, "capturing autonomous agent activity most tools can't see at all", and multi-touch attribution that "traces conversions across every touchpoint, including LLM and AI-referred traffic". Conversation intelligence links chatbot and voice interactions to on-site friction, and MCP connections bring experience data into tools including Claude, ChatGPT, Microsoft Copilot and Cursor.

Sense, its agentic layer, now covers experimentation: it can form hypotheses, design and code tests, launch them and validate revenue impact, and the release says "all you need to do is approve". CMSWire's coverage confirms the same feature set and the October 13 date. The platform is in beta for selected customers, with broader access rolling out in phases.

What it does not tell you

The release contains no percentage or outcome statistics on agent or LLM traffic. Its only quantitative claims are scale descriptors: "trillions of interactions" and ten years of zone-level data behind a new Benchmarks feature. A Škoda case study teaser elsewhere on the site cites 40% more test drive visits, but that figure is not in the release, so do not quote it as evidence for the agent-analytics claim.

That gap is the point. Nobody outside the vendors can currently tell you what share of your sessions are agents, whether they convert, or whether they behave like humans. Those are testable questions on your own data.

Why blended numbers mislead

A blended conversion rate is an average of two populations. If agent sessions browse heavily and rarely convert, they drag the average down and make a healthy page look broken. If AI-referred humans arrive with high intent, they inflate the average for a channel you did not buy. Either way, a decision made on the blended number is wrong for part of your traffic. That is a hypothesis, not a finding, and segmentation is what turns it into one.

It also changes how you report upward. A CMO who hears that organic is down deserves to know whether people stopped coming, or started arriving through an assistant and being credited somewhere else.

How to segment without a new platform

You do not need a platform migration to start. Three sources get most of the way there:

• Server and CDN logs. Filter user-agent strings that identify AI assistants and crawlers, such as ChatGPT-User or PerplexityBot, and keep them out of human session counts.
• Referrer data. Group visits arriving from AI assistants into one channel so they can be reviewed apart from generic referral traffic.
• Conversion paths. For each order or lead, record whether an AI-referred touch appeared before it converted.

Keep a human on the experiment gate

The approval step in agentic experimentation deserves scrutiny. The release describes the loop as end to end with approval, but does not say what the approval involves. An agent that writes and launches tests against your revenue pages is only as safe as the review attached to it. Define who approves, what a test may touch and which metric stops it before a vendor's default decides for you. The same rule applies to any agent you build yourself.

What to do about it

1. Segment agent and LLM sessions out of human sessions in your reporting this quarter, using logs and referrers.
2. Create an AI-referred channel group and review its conversion rate, assisted conversions and landing pages monthly.
3. Treat vendor claims without volume data as hypotheses. Ask any analytics vendor for agent-traffic share on a comparable site before paying for it.
4. Write an approval policy for any agent that launches tests: scope, approver and stop metric.
5. Re-baseline your headline KPIs once segmented, and tell leadership which numbers changed definition.