Audience Analysis

Overview

Audience Analysis answers questions about what an audience contains. Ask in plain language and it computes the answer over the actual rows. That can be how many match a condition, the breakdown of a column, the ten accounts highest on a score, or the themes that recur across a column of free text.

An audience is a working list of companies or contacts with columns attached. Some columns are structured: industry, employee range, signal scores. Others hold free text: research findings, descriptions, notes. Audience Analysis reads both. That matters because an audience of a few hundred rows is too big to read by hand, and the free-text columns most teams collect usually go unread. This tool turns them into something you can count, group, and rank.

What You'll Get

A written answer to the question you asked, computed over the audience's full contents rather than a sample, formatted as a list or table where that fits.

Depending on the question, the answer includes:

• Counts, totals, and averages, with any conditions you set applied
• Distributions and cross-tabs, with counts and percentages
• Ranked lists, such as the top ten accounts on a score
• Fill rates per column, showing where the data has gaps
• Free-text columns classified into categories you define, with the rows in each
• The supporting rows behind a number, when you ask for them

Best For

• Getting oriented in an audience you didn't build: which columns exist, how populated they are, what values they hold
• Breakdowns, rankings, and pivots that would otherwise mean exporting to a spreadsheet
• Turning free-text columns into counts: recurring themes, categories you name, rows whose text matches a condition you describe
• Classifying text and then pivoting the result against structured columns, such as grouping accounts by the pain point in their notes and comparing average scores across groups
• Data-quality checks before you act on an audience

To see example rows or preview specific columns, use Get Audience instead. To find the audience in the first place, start with Browse Audiences.

How It Works

Provide an audience and a question. There are no filter parameters to configure; the specificity lives in how the question is phrased. You are briefing an analyst.

The tool handles three kinds of question:

• Structured. Counts, distributions, cross-tabs, rankings, aggregations, and fill rates over the typed columns: picklists, numbers, dates, currency, booleans.
• Unstructured. Reading the free-text columns across rows: synthesizing recurring themes, sorting rows into categories you name, pulling out mentions of products or objections, and flagging rows whose text meets a condition you describe.
• Both at once. One call answers both kinds, so it can classify text and then pivot the result against structured columns. Ask which pain points cluster in which employee ranges, and the same question classifies the notes and breaks the groups down by the firmographic columns. A spreadsheet pivot has no way to do this, because it can't read the text.

Writing effective questions

• Name the columns, in quotes, exactly as they appear. Audiences often carry several columns that look similar.
• Supply the categories when you ask for a classification. Given none, the tool picks its own, and they may differ from the ones you'd use.
• Ask for the shape you want. Counts, percentages, a ranked list, a table: say so, and that is what comes back.
• Ask for the supporting rows. "How many, and which ones" costs nothing extra over "how many."
• Bundle related questions into one call, so the answer arrives as one coherent report. Very long multi-part questions can time out, though. If a report has many independent sections, split it in two.

A worked example

Take a validation run on a 148-company audience: software companies in the Boston metro, 500 or more employees. One question asked for fill rates per column, the employee and revenue distributions, the most frequent industries, and a classification of a custom free-text column. One report answered all of it.

The answer reported 100% fill on the firmographic columns and 34.5% on the custom text column, matching exactly what had been written into it. Employee counts split 78 rows at 500 to 1,000 and 56 at 1,000 to 5,000. Geography read 136 Massachusetts and 12 New Hampshire, correctly reflecting a metro filter rather than a state one. The free-text column came back sorted into the categories supplied in the question, with 12 problem rows flagged by name, each with a reason.

What to keep in mind

• Answers reflect the audience, and the audience can be wrong. Stale or misclassified source records are faithfully reported. That cuts both ways: the tool is good at surfacing them, as the flagged rows above show.
• Classification is built for triage. Recall is strong, but a flagged-row list can miss an edge case. Spot-check before treating one as an audit.
• List columns count as whole values. A column holding several values per row is tallied on the combined value. Counts are usable; a true per-value distribution over a list column is out of scope today.

Technical Details

Credits

AI action credits. Each question runs a sub-agent that computes over the audience's contents, so cost scales with the size of the audience and the complexity of the question. See Credits & Billing.

Requirements

An audience id, which the model takes from Browse Audiences, Get Audience, or Upsert Audience rather than guessing. Free-text analysis needs the text columns populated; if an AI column is still enriching, check its fill state with Get Audience first.

Automatic Tool Selection

The AI reaches for Audience Analysis when you ask "how many," "what's the breakdown of," "which rows rank highest," "segment these by," or "summarize this audience." For a preview of example rows it uses Get Audience instead, and when it only has an audience name it calls Browse Audiences first. When a request ends in an action, such as updating the rows a condition matches, it uses this tool to identify the rows and then hands off to Manage Audience Rows.

Architecture

The question is answered by a downstream analysis agent that sees only the audience and the question, without the surrounding conversation. So the question must be self-contained. Include exact column names, the categories to classify into, and every calculation the report needs, in one query. The agent computes over the audience's real contents rather than a sample, then returns a written answer sized for the model that asked, so a large audience never floods the primary context window.

Example Prompts

"Give me the distribution of the 'Industry' column with counts and percentages."

"How populated is each column in this audience? Show a fill rate for each."

"Which ten accounts score highest on 'Fit Score', and what do they have in common?"

"Classify each account's 'Research Notes' as evaluating, expanding, or at risk, and count each group."

"Group rows by the main pain point in their notes, and give me the average 'Intent Score' per group."