Testing ad creative on a small budget usually means running several short copy variants against the same audience to see what lands, rather than committing a whole budget to one version. This compares both tools on that one job: generating a batch of meaningfully different variants, not just reworded duplicates.
The test
We asked both tools for eight short ad copy variants (headline plus one line of body text) for the same product and audience, explicitly asking for variety in angle (one lifestyle-benefit angle, one problem-solution angle, one social-proof angle, one urgency angle) rather than eight versions of the same idea.
Where ChatGPT felt stronger
Defaulted to punchier, more ad-native phrasing without much extra prompting, closer to what you'd actually want to paste into a Facebook or TikTok ad without heavy editing. It also generated a wider spread across the requested angles on the first try.
Where Claude felt stronger
When asked to explain the reasoning behind each variant (why this angle, for this audience), Claude's explanations were more useful for deciding which variants were actually worth testing versus generating filler, helpful if you're trying to build judgment about what makes a good variant, not just get a batch of options.
Where both struggled equally
Neither tool reliably avoided generating at least one or two variants that were close rephrasings of each other despite the explicit instruction for variety. Expect to manually cut duplicates from either tool's batch rather than trusting the full set is meaningfully distinct.
Our take
For a fast batch of test-ready copy with less editing, ChatGPT was the slightly better starting point in our runs. If you want the reasoning alongside the copy (useful if you're newer to running ad tests and want to build that judgment), Claude's explanations added more value. Either way, plan to manually trim duplicate-feeling variants before anything goes into an actual test.
A prompt template for ad copy variants
This is the same brief we used in the test above, written out so you can reuse it with either tool. The angle split is what does most of the work: asking for "eight variants" with no structure is how you end up with eight rewordings of one idea.
Write 8 short ad copy variants for [PRODUCT NAME].
Each variant: one headline and one line of body text.
Product facts (use only these): [FACTS]
Audience: [WHO THEY ARE AND WHAT THEY CARE ABOUT]
Offer: [OFFER, OR "none"]
Platform: [e.g. Facebook / Instagram / TikTok]
Brand voice: [3 WORDS, e.g. warm, plain, a bit dry]
Words or claims to avoid: [LIST]
Spread the variants across these angles, two each:
- Lifestyle benefit
- Problem and solution
- Social proof (only use proof I've given you: [REVIEW QUOTES / CUSTOMER COUNT / "none"])
- Urgency (only if the offer above has a real deadline; otherwise write a second problem-and-solution variant instead)
The two variants within each angle must use different ideas, not reworded versions of each other. Label each variant with its angle.
After the list, for each variant, explain in one sentence why it might work for this audience and what we'd learn if it won.The last line asks for the reasoning that Claude handled well in our runs. With ChatGPT it's still useful, because it forces each variant to have a reason to exist. If you only want the copy, delete that line.
Because neither tool reliably avoided near-duplicates, a second pass helps narrow the batch down. It doesn't replace reading the list yourself, but it gives you a starting point:
Here are [NUMBER] ad variants for [PRODUCT NAME].
1. Group any variants that make the same core promise in different words.
2. For each group, say which one is strongest and why, in one sentence.
3. List any angle that is now missing or only covered once.
Variants:
[PASTE VARIANTS]How to choose which ad variants to test
Before anything goes live, run the shortlist through this checklist:
- Each variant differs in angle, not just wording. If you can't say in a few words what makes two variants different, test only one of them.
- Only the copy changes. Keep the image, audience and placement the same across the variants, or you won't know whether the copy made the difference.
- Every claim is true. Social proof has to come from real reviews or real numbers you can back up, and urgency needs a real deadline.
- It reads well on a phone. Check each variant in the ad platform's own preview, since a headline that works on a laptop can get cut short on mobile.
- Three or four at a time. As the FAQ below says, a small budget rarely gives a clean read on more than that.
- Write the hypothesis down. One line per variant ("busy parents respond to time saved more than to price") makes the result useful for the next round, win or lose.
Worked example: a batch for one product
Take a fictional bath accessories brand, "Tidewell", selling an extendable bamboo bath tray, as an example. The inputs: it holds a book, a phone and a drink above the water, it extends to fit different tub widths, and the audience is people who want a proper wind-down evening at home. There's no current offer and no review data yet.
A batch from the prompt above might include:
- Lifestyle: "Your Sunday night, upgraded." / "A bamboo tray for your book, your candle and your tea, right where you need them."
- Problem and solution: "No more soggy paperbacks." / "Keep your book dry and in reach while you soak."
- Social proof: "Loved by thousands of bathers." / "Join the people who've made bath time their favourite hour."
- Urgency: "Limited stock: order tonight." / "Don't miss your next quiet evening."
What a person changes before testing:
- The social proof line is invented. Tidewell has no review data yet, and the prompt said so. Cut both social proof variants.
- The urgency line is false. There's no stock limit or deadline, so "limited stock" goes too. A model will often reach for urgency language even when there's nothing to be urgent about.
- Two lifestyle variants are the same idea. "Your Sunday night, upgraded" and a second variant along the lines of "You deserve a proper break" both sell "treat yourself". Keep one.
- The final test set is three variants: one lifestyle, the soggy-paperback problem and solution, and a second problem-and-solution variant built on "phone, drink and book in one place". Same image, same audience, then let the results decide.
Common mistakes when generating ad copy with AI
- No angle structure. Without named angles, both tools tend to drift towards one idea phrased several ways.
- Testing everything you generated. Generating eight is cheap; spreading a small budget across eight isn't. Shortlist first.
- Letting AI supply the proof. Review counts, ratings and "best-selling" claims must come from you, not the model.
- Changing the image and the copy together. That's a new ad, not a copy test.
- Picking the cleverest line instead of the clearest. A variant that makes you smile isn't necessarily the one a stranger scrolling past will understand in a second.
- Starting from zero every round. Feed the winning variant and your notes back into the prompt, and ask for new variants on the winning angle plus one fresh angle.
FAQ
How many variants should a small budget actually test at once? Most small budgets don't have enough spend to get a statistically clean read on more than 3-4 variants at once. Generate more than that for options, but don't try to test all eight simultaneously.
Does this apply to video ad scripts too, not just text? We only tested short text copy here. Video ad scripts have their own pacing and hook considerations that a static text comparison doesn't capture. Treat this as a starting point for text ads specifically, not a video-script recommendation.
Is there a separate workflow for TikTok video scripts? Yes. Batch-writing TikTok Shop video scripts with AI covers a script skeleton, hook variants and where AI stops being useful for video.
Is ChatGPT or Claude better for writing Facebook ads? In our runs, ChatGPT gave punchier, more ad-ready copy with less editing, and Claude gave more useful reasoning about why each variant might work. Neither is a clear winner for every team. If you already pay for one, start there and use the prompt above with either.
How do I stop AI from writing near-duplicate ad variants? Name the angles in the prompt, ask for two different ideas per angle, then run a grouping pass like the second prompt above. Even then, expect to cut one or two by hand; neither tool got this fully right in our test.
Can AI predict which ad variant will win? No. The reasoning a model gives you is useful for deciding what's worth testing, but only a real test with real spend tells you what your audience responds to. Treat the explanations as hypotheses to check, not forecasts.
See ChatGPT and Claude directly, or AI-assisted email campaigns for a related copy-drafting workflow on the email side.
