import DefinitionCard from "@site/src/components/DefinitionCard"

# AI Marketing Copy Workflow

_Generate copy from a factual brief, then evaluate it against a specific message and audience._

## Introduction

AI can produce many copy variants quickly. Without a precise brief, those variants repeat familiar marketing patterns and make claims the product cannot support.

Use the model for drafting and revision. Keep positioning, factual approval, and experiment design with people who understand the product and audience.

## Understanding the Workflow

A **copy brief** defines the product truth, audience, message, voice, and desired action for one piece of copy.

**Product truth** is the evidence the copy may claim. It includes observed behavior, measured results, supported use cases, and approved limitations.

Copy quality and copy performance are different. A clear sentence can still address the wrong concern or make the reader doubt the claim. Performance requires a defined outcome and measured reader behavior.

## Applying It in Practice

Prepare a brief with:

- one audience and the problem they recognize
- one supported product claim
- evidence for that claim
- the page's place in the reader journey
- the action the reader should take
- voice examples and banned language
- legal or compliance constraints

Ask for variants that test distinct message hypotheses. Five rewrites of the same claim provide less information than two variants built around different reader concerns.

Evaluate each draft against the brief:

| Check | Question |
| --- | --- |
| Truth | Does every explicit and implied claim have evidence? |
| Audience | Does the copy address a concern this reader has? |
| Message | Is the main point clear after one reading? |
| Voice | Does it match approved examples without copying them? |
| Action | Does the next step fit the page and reader stage? |

Revise with concrete constraints. “Keep the measured deployment result, remove the unsupported cost claim, and shorten the heading to eight words” is testable feedback.

For an existing page, preserve proven structure unless the experiment is designed to test it. Changing the message, layout, and call to action at once makes results hard to interpret.

Measure variants through an agreed experiment or user-research method. A model cannot predict conversion from prose alone.

## Engineering Considerations

Do not ask the model to discover positioning from a feature list. Positioning requires customer evidence, competitive context, and a choice about which audience to serve.

Check implied claims as carefully as explicit ones. Words such as “automatic,” “secure,” and “instant” create expectations even when the copy gives no number.

Protect customer data and unreleased product details in prompts. Use approved tools and sanitize research material before sharing it with a model.

Human review is required for pricing, legal claims, regulated products, testimonials, and comparative statements.

## Scaling and Operations

Maintain briefs as versioned source material. Update them when product behavior, evidence, positioning, or voice changes. Do not reuse a campaign brief where the audience or desired action differs.

Limit generation to the team's evaluation and testing capacity. Unreviewed variants create inventory, not learning.

Record the hypothesis, approved copy, audience, dates, and outcome for each experiment. Feed measured results into later briefs without treating one result as a universal writing rule.

Keep a claim register for important numbers and comparative statements. Link each claim to its evidence, owner, and review date.

## Next Steps

- [AI Documentation Writing Workflow](./ai-documentation-writing): produce factual technical content from source material
- [AI Feature Development Workflow](./ai-feature-development): implement the product behavior behind the copy
- [Human-in-the-Loop Review Workflow](/learn/workflows/git/human-in-the-loop-review): require human approval for generated content
- [What is Human-in-the-Loop Development?](/learn/concepts/ai-engineering/human-in-the-loop-development): match oversight to risk