Ali Fakhar
فارسیGet in touch

How to generate and test ad creative variations with AI

  • Level: Intermediate
  • A few days to set up

Ali Fakhar · Updated: Sept 2026

The short answer

Write one hypothesis first, then ask a model for variants that change only that element. Review every claim, run the variants with a budget and a decision rule agreed before launch, and record what won and why, so the next test starts from the learning.

On this page

At a glance

The marketing problem
Producing more variants does not automatically produce better learning. Each creative needs a clear hypothesis, a controlled comparison and a decision about what to keep.
Who this is for
Paid-media teams on LinkedIn, Google, Meta and similar platforms.
Tools involved
  • LinkedIn Campaign Manager
  • Claude API
  • Google Ads

Where I have run it

478 ad creatives tested across 65 campaigns

View case study ↗
Counted in LinkedIn Campaign Manager across the full programme. The same run cut blended cost per lead by 85%, first full quarter against the most recent one.

What you need

Before you start

  • One test hypothesis and the audience it applies to
  • Approved claims and proof points
  • A measurement plan: the metric, the minimum sample and the decision rule

What goes in

  • The creative brief and the offer
  • Your voice guide and past approved ads
  • Platform limits: character counts and formats
  • Results from previous tests

How it works

  1. 01 · You decideWrite one hypothesis
  2. 02 · AI stepGenerate constrained variants
  3. 03 · ToolRun the planned test
  4. 04 · You decideReview the result
  5. 05 · OutputRecord the learning
  • Tool
  • AI step
  • You decide
  • Output

Step by step

  1. 01 You decide

    Write one hypothesis

    Choose the audience and the single difference to test, such as the problem the customer names against the product benefit. Write down the metric and the decision rule now.

  2. 02 AI step

    Generate constrained variants

    Ask for variants that change only the element under test, using approved claims only.

    Prompt to copy

    We are testing this hypothesis: [hypothesis]. Write [number] ad variants for [platform] for [audience]. Change only [the element being tested]; keep the offer, call to action and tone the same. Respect these limits: headline [x] characters, body [y] characters. Use only these approved claims: [paste]. Label each variant with the angle it tests.

  3. 03 Tool

    Run the planned test

    Launch the variants with equal budgets in the same campaign and audience, using the platform’s experiment feature where there is one.

  4. 04 You decide

    Review the result

    Once each variant reaches the agreed sample, check how spend was distributed and the quality of the leads before deciding what to keep.

  5. 05 Output

    Record the learning

    Save the hypothesis, the variants, the results and the decision together, so the next test builds on them.

Where a person decides

A marketer approves every variant’s claims before launch and decides what to keep once the agreed sample is reached.

What to watch for

  • More variants split the budget and can leave every one short of a usable sample.
  • Changing several things at once means you cannot tell what worked.
  • Platforms deliver ads unevenly, so check spend and impressions per variant.

How to tell it is working

  • The decision rule was written before the test started.
  • Each variant reached the minimum sample.
  • Lead quality is checked, not only click-through rate or cost per lead.

An illustrative example

Input

Illustrative: test whether naming a specific customer problem produces more qualified enquiries than naming a general product benefit.

Output

Two reviewed copy directions with the same offer and audience, four variants each, plus the hypothesis, the metric and the rule for deciding.

Want this running in your team?

AI workflow implementation. Turn a repetitive marketing task into a workflow your team can use with confidence.

View service
All AI in marketing workflows ↗
Analytics preferences

Current preference: off