> ## Documentation Index
> Fetch the complete documentation index at: https://docs.permutive.com/llms.txt
> Use this file to discover all available pages before exploring further.

# 1.2 - Audience Extension, Comparison & Detailed Analysis

> Explain an audience, not just size it

<Note>
  **Your agent can now explain an audience, not just size it**

  Adds an analysis layer: overlap, index, incremental reach, engagement and
  trends over time. Compare an audience against benchmarks, and discover cohorts
  that would extend it by affinity on indexing and overlap rather than semantic
  matching alone.

  **Status:** Closed testing · Invitation only · Tool interfaces may change
</Note>

## New capabilities

### Make a pitch defensible, not just responsive

> *"Measure this audience against the cohorts this advertiser already buys, and
> against our premium titles, so I can see overlap and index for each. Restrict to
> the brief's market."*

An agent can compare an audience against benchmark cohorts you choose - the
advertiser's existing buys, a category set, your premium inventory. For each one
it returns:

* **Overlap**: How many of the audience's users that cohort shares
* **Index**: How much more likely an audience member is to be in that cohort than the average user

This is what turns *"this audience is large"* into *"this audience over-indexes
2.4× on luxury automotive intent and overlaps only 12% with what this advertiser
already buys."*

### Find the cohorts you would never have searched for

> *"Find cohorts that would extend this audience based on index and overlap rather
> than just theme or keywords. Give me stats on how these could improve the
> audience to meet my campaign goals"*

Given a seed audience, an agent can search the whole cohort library for the
cohorts whose users index well against it and provide a good extension into a
different user base. You do not name the candidates - the tool finds them.

This routinely surfaces results that may not be intuitive, but can provide
meaningful improvements to plans and campaigns. It complements the existing
ability for agents to search for cohorts based on theme, keywords or audience
descriptions, allowing the agent to have multiple approaches for extending
audiences

### Size a change before making it

> *"For any cohort you recommend adding, measure the audience I would end up with
> and tell me the incremental reach"*

Every suggestion carries its **incremental reach** - the net-new users it would
add, rather than its standalone size. A large cohort that mostly duplicates what
you already have adds very little, and that difference is what makes a
recommendation worth acting on.

Near-duplicates of the seed are filtered out, so a cohort that merely re-surfaces
the audience you already have does not crowd out genuine extensions.

### Separate redundancy from precision

> *"Show me what each cohort contributes on top of the others and tell me if any
> are redundant. For anything heavily duplicated, tell me whether it carries a
> sharper signal worth keeping as a precision layer."*

An agent can show which cohorts are carrying the audience and which are
duplicated by the others.

Redundant on reach does not mean worthless. A heavily duplicated cohort carrying
a sharper intent signal can come out of the reach build and go back in as a
precision layer on its own line. This step also tends to decide the pitch:
whether to lead with the broad audience and treat the brief's signals as
qualifiers, or to argue that the intersection is the product.

### Talk about engagement and momentum, not just size

> *"Which of my interest cohorts show significant engagement?"*

Audience measurement now returns engagement alongside reach, including:

* **Engaged time**: Total and average time the audience spends engaged
* **Views per user**: Average page views per user

### Find trends in audience performance

> *"How are my 'news' cohorts performing this month, are there any big risers or
> fallers?"*

Audience measurement now goes one step further and can break down reach
day-by-day, giving your agent the power to spot trends in behaviour.

## How it works

Two new read-only tools, plus an extension to audience measurement:

| Tool                                                            | What it does                                                                                                                                                                |
| --------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| [`compare_audience`](/api/mcp/compare_audience)                 | Compares an audience against specific cohorts - each one's shared users, overlap share, and index and the incremental reach                                                 |
| [`find_audience_extensions`](/api/mcp/find_audience_extensions) | Analyses your whole cohort library to find cohorts that would **extend** a seed audience, ranked by index, each with its overlap, incremental reach and a confidence level. |
| [`measure_audiences`](/api/mcp/measure_audiences)               | Updated to add engagement metrics, daily and weekly time-series, and a variable date window.                                                                                |

<Info>
  **Which tool will the agent use?**

  The two new tools are siblings, and the difference is who picks the cohorts.
  With [`compare_audience`](/api/mcp/compare_audience) the agent must specify them

  * use it when you already know the benchmark set, such as the advertiser's
    current buys or a tagged category. With
    [`find_audience_extensions`](/api/mcp/find_audience_extensions) **the tool**
    finds them - use it when you want discovery.

  Both report index against the baseline population of the workspaces you ask for,
  so figures stay comparable across calls as long as you keep the same workspaces.
</Info>

Some figures are estimated from sampled data, so accuracy can vary, but this is
highlighted to the response to the agent, and once your agent has identified
candidate cohorts you should expect it to explicitly confirm sizing. Every
suggestion therefore carries a **confidence level**, and low-confidence results
are excluded unless you ask for them. Where a seed audience overlaps too little
with the library to extend from at all, the agent is told so explicitly rather
than being handed a thin list of noise.

For a full tool reference, see the [Permutive MCP tool documentation](/api/mcp/tools)