AI guide

Test ad copy with AI: what ChatGPT and Claude can tell you and what an audience adds

Testing ad copy with AI can mean three things: asking an assistant like ChatGPT or Claude, letting a simulated audience read it, or running variants live. Each answers a different question. Here is which one answers what, and how to get more out of each.

Start for free 350 credits to start, no payment details

Last updated:

Three ways to test ad copy with AI

MethodAnswersDoesn’t answerNeeds
Ask an AI assistant (ChatGPT, Claude)Is the copy clear? What alternatives are there?Whether your audience prefers itOnly the copy
Simulated audienceWhich version the personas prefer and what holds them backWhether the version gets more clicks in the marketPersonas based on real customer voices
Live test with adsClicks and inquiriesWhy a version winsBudget, traffic and run time

The three complement each other: write with the assistant, shortlist with the audience, and test the two or three strongest live. For a comparison of the methods, see message testing.

How to get more out of ChatGPT or Claude

  • Give context: audience, channel and where the reader comes from. Copy for a search ad has a different job than copy for a newsletter.
  • Compare instead of scoring: put two versions side by side and ask about the difference. Scores from 1 to 10 tell you little.
  • Ask for the strongest objection: what stops a skeptical reader from clicking?
  • Plan for agreement: Language models tend to agree with their users. A study by Anthropic calls this “sycophancy” and finds it in several widely used assistants. So ask for problems rather than approval, and check whether the criticism still comes when you don’t ask for it.
  • Don’t use numbers you can’t back up: assistants like to suggest specific percentages. If a number isn’t in your own material, it doesn’t belong in the ad.

Three questions for every piece of ad copy

  • Would someone with no background knowledge understand what is on offer?
  • Does the copy match what the reader just searched for or saw?
  • Does it say how the result comes about, not just that it does?

For headlines in particular: headline analyzer. For buttons: CTA text.

Test ad copy with 99takes in Claude

99takes writes challenger versions from your material, with no unsupported numbers. Three versions of a headline cost about 6 credits (challenger versions). If you already have several drafts, a shortlist pre-sorts them for 3 to 5 credits.

The best version then competes with your current one, read by personas from your audience. With 12 personas that costs about 73 credits, with each persona also reading it with the other personas’ jobs in mind; one-to-one, about 19.

The claim check shows whether competitors already make your core claim, for 3 to 7 credits.

All of this is a simulation before launch, not a market measurement. It helps you go into the live test with the stronger versions.

Frequently asked questions

Can ChatGPT test ad copy?

ChatGPT can rate copy, find unclear passages and write variants. It doesn’t measure behavior, and without details about your audience it judges from general knowledge. To find out which version your audience prefers, you need readers from that audience, real or simulated.

Test ad copy with AI or run an A/B test?

Both, one after the other. Use AI to shortlist while many versions are still in the running and you don’t want to spend budget yet. In the end, an A/B test with real ads decides what brings more clicks and inquiries.

What is a simulated audience?

Personas played by a language model that are meant to read copy the way people from your audience would. How well that works depends on what they are based on.

What does it cost to test ad copy with 99takes?

Writing three versions of a headline costs about 6 credits, a shortlist 3 to 5, a comparison with 12 personas about 73. You start with 350 credits, no payment details.

Related reading

Shortlist ad copy before you spend the budget

Create an account, connect 99takes to Claude and have your audience read two versions side by side. You see the price before anything starts.

Start for free

Sources (retrieved )

  1. Sharma u. a. (Anthropic): Towards Understanding Sycophancy in Language Models, 2023
  2. UX studio: Can AI take over usability testing?