A/B Test Ad Copy Variations

Create distinct A/B ad variations by changing only a single element.

Prompt · 3 variables

I want to A/B test ad copy for {{media distribution channel}}. The hypothesis I want to test is: "{{hypothesis to test}}".

Based on the original text below, please create Option A and Option B. Keep Option A almost identical to the original with minor polishing only. For Option B, change only the single element related to the hypothesis. Keep all other sentences, length, and tone identical to Option A.

Present the results in a table. The columns in order should be: "Variant | Copy | Modified Element | Expected Reaction". Below the table, specify a single metric to evaluate the winner of this test.

Keep the character count difference between both options within 10% of the original text. If the original is very short and 10% is less than 3 characters, a difference of up to 5 characters is acceptable. Do not use unsubstantiated claims like "#1", "Lowest Price", or "100% Guaranteed", and keep the number of emojis identical in both options.

After writing the response, compare Option A and Option B side by side to perform a self-check on the actual differences. If more than one element was changed, revise it and display only the final version.

Original text: """ {{original phrasing}} """

Copy, then paste here · ChatGPT and Claude open with the prompt filled in Open in ChatGPT ↗Open in Claude ↗Open in Gemini ↗ Edit in builder Download classroom card

Why it is written this way

Context
I want to A/B test ad copy for {{media distribution channel}}. The hypothesis I want to test is: "{{hypothesis to test}}".
Task
Based on the original text below, please create Option A and Option B. Keep Option A almost identical to the original with minor polishing only. For Option B, change only the single element related to the hypothesis. Keep all other sentences, length, and tone identical to Option A.
Format
Present the results in a table. The columns in order should be: "Variant | Copy | Modified Element | Expected Reaction". Below the table, specify a single metric to evaluate the winner of this test.
Constraints
Keep the character count difference between both options within 10% of the original text. If the original is very short and 10% is less than 3 characters, a difference of up to 5 characters is acceptable. Do not use unsubstantiated claims like "#1", "Lowest Price", or "100% Guaranteed", and keep the number of emojis identical in both options.
Self-check
After writing the response, compare Option A and Option B side by side to perform a self-check on the actual differences. If more than one element was changed, revise it and display only the final version.
Input
Original text: """ {{original phrasing}} """

When you ask AI to generate ad copy for A/B testing, it easily provides "two versions." The issue is that both versions often differ simultaneously across headlines, length, emojis, and benefit framing. Even if Variant B wins, you cannot identify what drove the win, leaving no actionable insights for future tests. This prompt focuses entirely on isolating a single variable.

In the Task paragraph, anchoring Option A to the original text is essential. If both options are rewritten from scratch, the baseline disappears. Locking Option A ensures that any change in Option B represents a measurable difference.

In the Format, the "Modified Element" column forces the model to state explicitly what it altered. If it lists multiple items, the violation is immediately obvious in the table. The Constraint regarding character count and emoji parity controls subtle variables that skew click-through rates.

Finally, the last paragraph instructs the model to self-check and refine its output. The most common pitfall in AI-generated split testing is altering multiple elements; demanding a side-by-side comparison filters out most multi-variable errors before output. Restricting the evaluation to a single metric serves the same goal: evaluating CTR, conversion rate, and saves at once leads to ambiguous interpretations. Deciding the success metric beforehand ensures clean testing.

Unfamiliar terms? See Aha AI: output-format, prompt

Compared with a bad example

Common bad example

Write 2 versions of this ad copy for an A/B test.

(Paste original copy)

This produces two completely different ad copies—for instance, one short and emotional, the other long and benefit-driven. If one wins, you cannot tell whether brevity or emotional tone caused the lift, leaving no reusable insight for the next campaign.

Variations

When testing multiple elements sequentially

When testing multiple elements sequentially

I want to run a multi-stage test on ad copy for {{media distribution channel}}. The hypothesis is: "{{hypothesis to test}}".

Create 3 options by modifying only one specific part from the original below. Option 1 should change only the opening sentence, Option 2 only the benefit phrasing, and Option 3 only the final call-to-action, keeping everything else identical to the original. Present the results in a table with columns: "Modified Part | Before | After".

""" {{original phrasing}} """

Predefining the test positions allows for phased experimentation. Each of the three options touches only one specific area.

When auditing existing A/B test copies

When auditing existing A/B test copies

Below are two A/B test copy drafts I wrote for {{media distribution channel}}. The hypothesis is: "{{hypothesis to test}}".

Do not rewrite them; please audit them instead: ① List all actual differences between the two versions, ② Identify any differences unrelated to the hypothesis, and ③ Explain why interpreting test results might be difficult if run as-is. Finally, provide revised versions that eliminate only the unrelated differences.

""" {{original phrasing}} """

Human-written drafts often contain multiple unintended differences. Use this to audit and isolate variables before launching a test.

Model notes

If multiple elements differ between the two options, prompt again with: "Keep only one modified element in Option B and use the exact same wording from Option A for the rest."

If the model changes multiple elements, follow up with: "Keep only one modified element in Option B and match the rest of the text word-for-word with Option A." Because LLMs struggle with exact character counting, verify the final length manually.

Related prompts

Last updated 2026-09-02 · Found a mistake? Let us know