# Measure domain

> You see what agents do with you: visits, questions, actions, failed tasks and the outcomes they lead to, counted the same way everywhere.

Source: https://agentrelations.org/framework/domains/measure

## The outcome and why it matters

Without measurement, Agent Relations is a set of hunches. Measure tells you what agents do with you, whether they succeed, what they say about you and what it is worth. It also turns the other five domains into a managed program. The Institute defines the core metrics below so that figures mean the same thing from one organization to the next and can be compared in the Index.

An analogy

Think of an annual report. Before it existed, each company described its health in its own words. Standard definitions let readers compare, and let the company see itself clearly.

Led by: Agent Relations practitioners. See the [responsibility matrix](https://agentrelations.org/framework/personas).

## Capabilities

5 things an organization does in this domain. Standards named here are explained in the Standards Map.

Capability 1

### Agent traffic classification

Separating agents from people and from other automated traffic, and sorting agents by what they are doing: training crawl, search indexing or fetching for a person’s request.

What good looks like

Traffic is classified by verified identity where possible and by declared user agent otherwise, and the method is documented. Unidentified automation is reported separately, never counted as agents.

Typical standards and techniques

- Web server and CDN logs
- Web Bot Auth and verified-bot signals
- Analytics referral data (for example the utm_source value some assistants add to links)

Capability 2

### Interaction logging

A record of the questions agents ask and the actions they take, with the result of each.

What good looks like

Each interaction has an identifier, an agent identity, a time, an outcome and a reason code for failure. Logs are retained under a stated policy and protect personal data.

Typical standards and techniques

- Structured logs and trace identifiers
- OpenTelemetry for tracing
- Reason-code taxonomies for failures

Capability 3

### Answer quality testing

Checking that answers, yours and those agents give about you, are accurate.

What good looks like

A fixed set of test questions is run each month against your endpoint and against major assistants, and results are scored against the source of truth by a reviewer.

Typical standards and techniques

- A maintained prompt panel
- Sampling and scoring protocols
- Source-of-truth fact register from Describe

Capability 4

### Outcome attribution

Tracing orders, bookings, leads and applications back to the agents that produced them.

What good looks like

Attribution uses signed identity or a declared referral first, then assisted-outcome estimates labeled as estimates. The method is stated beside every figure.

Typical standards and techniques

- Referral parameters
- Order-source fields in commerce systems
- Agent identity from Trust

Capability 5

### Reporting and benchmarking

A regular report with the same definitions every period, and comparison against peers.

What good looks like

A monthly dashboard shows the core metrics with trends. A quarterly summary goes to executives. Definitions are versioned.

Typical standards and techniques

- The core metrics defined in this domain
- The Agent Relations Index for peer comparison
- Versioned metric definitions

See the [Standards Map](https://agentrelations.org/standards) for who owns each standard and how mature it is. To learn the skills, see the [Specialist certification](https://agentrelations.org/learn/certification/specialist) (Marketing and Engineering).

## KPIs

These are the core Agent Relations metrics. The Institute defines them so that the numbers mean the same thing everywhere.

Agent visits

The number of sessions by identified agents on your public pages and endpoints in the period. A session is a run of requests from one agent identity with no gap longer than 30 minutes. Report by agent and by purpose (training, search indexing or user-requested fetch). Exclude people and unidentified automation.

Agent questions answered

The number of distinct agent questions received through your question endpoints or tools that got a substantive answer, shown beside the number received. Errors, timeouts and bare "no data" replies are not answers.

Answer accuracy

The share of sampled answers judged correct against your source of truth by a human reviewer, from a monthly sample of at least 100 answers. Measure two views: the answers you give agents, and what agents say about you on your prompt panel.

Agent-initiated actions

The number of actions (orders, bookings, applications, quote requests) that an agent started through your callable actions, with the completion rate: completed divided by started.

Failed agent tasks

The number of agent tasks that ended without the intended result: an error, a timeout, a rejection or an abandonment. Report the failure rate (failed divided by attempted) and the count by cause, so each cause can be fixed.

Agent-referred outcomes

The count and value of outcomes (sales, bookings, leads, applications) traced to an agent by signed identity or declared referral. State the attribution method, and never count one outcome as both agent-referred and person-referred.

Share of agent recommendations

Across a fixed panel of prompts a real customer would ask, the percentage of agent responses that recommend or name you, out of all responses. Run each prompt several times, on the same agents each month, and report by agent beside named competitors.

How the core metrics work together

Visits show reach, questions answered and accuracy show quality, actions and failed tasks show whether agents can get things done, referred outcomes show value and share of recommendations shows how agents rank you. Read them together: never report visits alone.

## What each maturity level looks like here

1. 1
Level 0: Closed
No agent traffic is separated from other traffic. Nobody knows what agents do.
1. 2
Level 1: Visible
Agent visits are counted from logs. Nothing is reported beyond that.
1. 3
Level 2: Described
Visits, questions and the first failures are counted with documented definitions. Accuracy is tested occasionally.
1. 4
Level 3: Conversational
A monthly report covers the core metrics, including failed tasks and accuracy from a regular sample.
1. 5
Level 4: Callable
Outcomes are attributed to agents and share of recommendations is tracked on a fixed panel. Executives see a quarterly summary.
1. 6
Level 5: Trusted
Metrics steer investment, definitions are versioned and the figures are benchmarked against peers.

[See all domains side by side](https://agentrelations.org/framework/maturity) or take the [self-assessment](https://agentrelations.org/framework/maturity#self-assessment).
