From UX Metrics to AX Metrics: How Do You Measure Experience in the Agent Era?

Three shifts: from click to citation, from citation to action

Since the beginning of the commercial internet, digital marketing measurement rested on a single assumption: the user comes to your site, does something there, and we measure it. That assumption broke in two steps, and the third is happening now.

The first shift: SEO to GEO. As traditional search volume declined, a meaningful share of users began consuming the answer inside an AI interface without ever reaching your site. In that world success stopped being the user's visit and became the AI citing you as a source in its answer. Experience measurement moved from how a person behaves on the site to how a machine makes sense of it. Last year's article described that shift.

The second shift: citation to action. AI agents left the chat window. Agent-driven traffic grew sharply over the past year, and in B2B a meaningful share of purchase research now runs straight through an assistant. The agent no longer merely mentions you; it enters your site, compares products, fills the quote form, adds to basket. Being cited is necessary but not sufficient: an agent can cite you as a source and then complete the task at your competitor, because the form on your site cannot be filled.

The third shift: measurement splitting into three layers. Human experience metrics keep their validity; people keep coming. AI visibility metrics were added on top. Now a third layer is being added: agent action metrics. A brand's experience performance is now understood only by reading all three together.

The name for this change is Agent Experience (AX): how easily, safely and reliably an AI agent can use a website, a piece of software or an API. The third layer of the measurement frame measures exactly that.

The evolution of metrics: old indicator, new counterpart

Last year we said "citation value instead of bounce rate". This year we widen the table: every classic UX metric has a visibility counterpart and an agent action counterpart.

Three relationships in this table matter most.

Bounce rate and access success. Last year we redefined bounce as a "verification failure": the AI crawls you but leaves you out of its answer because it does not find you credible enough. In the agent era you have to go one step further back: can the AI crawl you at all? If OAI-SearchBot is blocked in robots.txt, if the content sits behind JavaScript, if the firewall silently rejects ClaudeBot, verification never even begins. The status codes agent user agents receive in your server logs are today's most basic experience metric.

Share of answer and task completion. A brand with high share of answer and low task completion is the most common profile of the agent era. The agent recommends you, the user says "fine, buy it there", the agent enters your site and gets stuck on a fake <div> button. The visibility investment is lost in the action layer. A team that does not read the two metrics side by side cannot see that leak.

Freshness and access latency. Crawl-based engines clearly prefer current data; content not refreshed quarterly drops out of citation. On the agent side latency means something different: while completing a task an agent makes dozens of requests within seconds, and every JavaScript render, every redirect, every slow API response raises the chance the task is left half done. Server-side rendering is far more critical than it ever was for classic search.

The three-layer measurement frame

At Webtures we now build experience measurement as a three-layer frame rather than a single dashboard. The layers do not replace one another; each measures a different user's experience.

Layer 1: human experience

Sources: Google Analytics 4, Microsoft Clarity, user testing. Metrics: conversion, exit rate, dwell time, heat and scroll maps, form abandonment, task success. Its question: what does the human visitor do on the site, and where do they struggle?

This layer has lost no value. People keep coming to your site, and more and more of them arrive on an agent's recommendation. A user brought by an agent is a high-intent visitor who has already made most of the decision; a poor experience on the site costs twice as much.

Layer 2: AI visibility

Sources: Brantial, regular engine queries with a defined prompt set. Metrics: citation rate, share of answer, visibility per engine, sentiment, brand accuracy, freshness. Its question: how often, how accurately and in what context do AI engines recommend me to my customer?

The critical property of this layer is multi-engine monitoring. Source overlap between engines is low; a brand watching a single assistant never notices its invisibility in the others. The prompt set must stay fixed so that monthly comparison means something.

Layer 3: agent action

Sources: server logs, agent task simulations, the AI Agent Readiness scan, MCP and API logs. Metrics: access success, task completion rate, break-off step, access latency, autonomous resolution rate, agent-readiness score. Its question: can the agent finish its job on my site, and if not, where and why does it break off?

This layer is the newest and the least built. Analytics does not see the agent; Clarity does not record it. The data comes only from server logs, simulations and external scanning. We described this layer's analysis method in detail in our article on agent trace maps.

The three layers together

The three layers should sit side by side on one management dashboard, because the real insight appears in the gap between them.

• Layer 2 high, layer 3 low: you are visible but nothing can be transacted. Priority: the access and tools layer.
• Layer 3 high, layer 2 low: everything works on the site but agents do not recommend you. Priority: context and authority.
• Layer 1 falling while layer 2 rises: traffic is down but visibility is up. That is not a loss but a channel shift, and it should be explained to leadership that way.

Where the classic UX audit stands in the Agent Experience era

Experience measurement on the human side has not lost its value; its scope has widened. We wrote that sentence last year too, and we defend it more strongly this year, because the outputs of a classic UX audit have become the direct input of the agent layer.

A UX audit systematically identifies a digital product's shortcomings in usability, accessibility and business goals. Its business value gathers at three points: an impartial outside view that breaks tunnel vision, comparison against competitors, and catching faults before they turn into development cost. All three still hold. What is new is that the audit methods have gained a third function.

Heuristic evaluation and user testing. They measure how intuitive the interface is for a person. Clear information architecture feeds machine readability. New: testing the same interface with an agent task simulation reveals elements that are intuitive for a person but unreadable for an agent, such as a button that is visually strong and semantically weak.

Analytical data analysis. Session recordings and heatmaps show which content meets a real need. New: that data is cross-read with agent log analysis. If the element people click most is the same element the agent cannot find, the problem is in the code.

Content and accessibility audit. Semantic HTML, heading hierarchy, descriptive copy and ARIA. Last year that was the content comprehension layer for language models. This year it is more: the accessibility tree is the main channel through which agents see the site. WCAG compliance has become a precondition of agent compliance. A button that is right for a screen reader is right for an agent.

In short, the classic UX audit was the infrastructure of the GEO strategy; now it is the infrastructure of the AX strategy too. Every experience simplified for a person produces a structure that is more readable and more usable for an agent. The point is not to drop the audit but to connect its outputs to all three layers.

Tying measurement to action: structured data, protocol, authority

A measurement frame stays a report unless it is tied to an action frame. Each of the three layers has its own area of fixes.

For visibility: machine-friendly content and entity authority. Direct answers: content should answer likely questions plainly in the first paragraph. Authority building: digital PR not only for backlinks but so the brand enters language models' knowledge base as a trusted data source. Entity optimisation: a consistent definition of brand, product and service in the knowledge graph through Organization, Product and Service schemas with @id and sameAs links. Schema alone is not enough; most AI crawlers read the visible HTML too, and the two must match.

For action: access, tools, protocol. The agent layer's fixes are technical but ordered. Access first: robots.txt and the firewall, server-side rendering. Then tools: semantic buttons, accessible names, consistent labels, machine-readable form results. Then protocol: an MCP server, WebMCP tool registration, and for commerce an ACP and UCP compatible feed with AP2 readiness, chosen by business profile. Investment that skips this order is wasted; a flawless MCP server on a bot-blocked site reaches nobody. For the detail of the protocol layer, see Agentic Commerce Readiness.

For both: trust and governance. An agent transacting with you raises the question of responsibility. Which action may the agent take alone, where is human approval needed, where is the transaction trail kept? A meaningful share of customers remains cautious about fully automated service; trust is a metric to measure and a layer to design.

The measurement and action loop should run monthly: measure the three layers, find the widest gap between them, fix that layer, measure again. Deterministic scanning tools make this loop comparable; because the same input gives the same score, progress is visible beyond argument.

From invisibility to reference, from reference to a transactable source

We closed last year's article like this: the question to ask is not "how many people clicked my site" but "how often do AI assistants recommend me to my customers". That question still holds, and a second one has joined it: after recommending me, can they complete the transaction with me?

User experience no longer lives behind the screen but at the heart of the algorithm, and increasingly in the agent's action. SEO listed you in search results. GEO made you the answer itself. Agent Experience makes you the address of the action that follows the answer.

At Webtures we build experience measurement across these three layers, read all three on one dashboard and turn the gap between them into action. To see where your site stands on the third layer, start with a free agent-readiness scan, and we can build the three-layer measurement programme together on the Agent Experience side.