Understanding Campaign Optimizer Agent concepts

Before we look at how tests run, it helps to learn a few terms the Campaign Optimizer Agent uses. These concepts—control, variant, tournament, and scoring—make it easier to follow what’s happening and interpret results.

Key terms

• Control — the current live message.
• Variant — an alternative, either generated or authored.
• Tournament — a head-to-head ranking of multiple variants generated using different strategies (prompts). A judging layer pits these variants against each other, and the top-ranked ones are suggested as starting points for A/B/C testing on the Review & Approve screen.
• Test — an A/B (or multi-variant) experiment you approve and launch.
• Lift — observed improvement vs. control during testing.

Tournament overview

The Campaign Optimizer Agent compares your control with multiple variants designed to improve a selected goal (for example, clicks or conversions). Multiple variants may be generated, but nothing goes live automatically—a test starts only after you approve a variant to run against the control.

A tournament is a simulated set of head-to-head comparisons among generated variants (not against the control). The judging layer ranks them by predicted effectiveness, and the top variants are surfaced as suggested options to start a multi-variant test.

To view tournament history: in the left navigation panel, go to Customer AI > Agents > Campaign Optimizer > Review & Approve > Content Generation (tab). Or select an opportunity, then open the Content Generation tab and check the Tournament history.

In the tournament view, each card shows the content and the reasoning for its rank. Select Show matches on a card to see its head-to-head results.

Ranking at a glance

• Variants appear as 1st, 2nd, 3rd, etc., with per-model scores (e.g., Gemini, Anthropic, OpenAI) and an overall Average on each card.
• The list is ordered by the Average score (highest first). For example, a variant with Average 1162 appears above one with 1094.
• The tournament is advisory—you choose which variant to approve for a live test.

“Show matches” details

Select Show matches on a card to view its head-to-head history. You’ll see:

• Outcome — e.g., Won vs Opponent (1030 ELO).
• Opponent content — the competing option it faced.
• Judge’s reasoning — a short explanation of why the winner is predicted to perform better (clarity, urgency, relevance, etc.).

Inside the tournament’s decision-making

Judge’s reasoning may reference familiar copy frameworks such as AIDA (Attention, Interest, Desire, Action) or RISE (Relevance, Impact, Specificity, Emotion). These callouts highlight the strengths the model predicts in one option over another.

• AIDA: How well the content grabs attention, sustains interest, builds desire, and prompts action.
• RISE: How the content scores on relevance, overall impact, specificity, and emotional resonance.

Optimization stats panel (right side)

The side panel summarizes performance by model and by strategy used to generate variants.

Model performance

Shows how different AI models performed when generating and judging candidate content.

Strategy performance

Shows how the generation strategies performed. Examples: Baseline (neutral/standard messaging), Signal to Noise (clarity-first), Touch of Intrigue (introduces light curiosity). These labels help you compare tone and approach.

Statistics on display

• Average ELO — the mean tournament score for a model/strategy across its generated variants.
• Variants generated — how many candidates it produced.
• Validity rate — the percentage of generated variants that appear in the tournament results.
• Best ELO — the highest single ELO achieved by any variant from that model/strategy.

  Try a mock tournament

Preview a tournament without affecting anything live. This mock run doesn’t start a real test and won’t change any in-progress results.

From an opportunity, open the Content generation tab.

Select Generate content to run the preview with your content. You’ll see a simulated set of results for demonstration.