Ad concepts vs. ad variations: how to plan an AI test
Separate new buyer arguments from execution changes, then record each source, creator role, and approval before generating ads.
An ad concept changes the buyer question or main product argument; a variation changes how the same argument is executed, such as its opening shot or crop. This is a planning distinction, not a promise about how a platform labels the asset. Record the source, claim, creator role, and approval needed before generating either.
Set the question and scope before generating
Write the concern, product answer and evidence in one sentence. Name the creator’s role and campaign context. Before putting a person’s recordings, images or voice into a tool, check that the campaign scope covers the material and ask what the tool stores or uses for training. Record the tool and account used so the team can check its current settings and terms.
Fill a concept and production brief
Use the same fields for each proposed concept:
| Field | What the team records |
|---|---|
| Viewer question | The concern this ad should answer |
| Main argument | Why the product addresses that concern |
| Evidence | The footage, measurements or supported explanation needed |
| Creator role | What the creator can show or say in the agreed context |
| Source material and permission | The files used and the scope that covers them |
| AI prompt and checks | The allowed facts, requested output and checks still needed |
| Expected response | What the team expects the viewer to understand or do |
| Difference from current ads | The changed argument, not just the changed style |
Reject a concept if its evidence is missing or the creator role asks for an experience the creator did not have.
Completed hypothetical brief-to-test matrix
| ID and type | Question and argument | Claim source | Creator role | Creator approval | Test note |
|---|---|---|---|---|---|
| C-01, concept | Will this meal container fit in a work bag? Show its measured size beside one named bag. | P-01: hypothetical product sheet with dimensions, to be checked against the real item before use. | Demonstrate the measured item and bag; do not claim it fits every bag. | Pending review of the final script and render. | Compare this full concept with the current ad; a result applies to the whole package. |
| V-01, execution of C-01 | Same buyer question and argument; open on a packing demonstration instead of a size card. | Same P-01; no new product claim. | Same role and evidence. | Approve this exact final version; approval of C-01 does not automatically cover the new render. | If testing the opening, keep other settings stable. |
Use the plain-text ad-test worksheet to record the comparison. This is a hypothetical example; confirm product facts and creator permission before production.
Write the prompt from approved facts
Treat the prompt as a production instruction, not proof that a claim is true. For example:
Using only these verified product facts: [facts]. Draft a 15-second ad for [viewer question] and [audience]. Creator role: [approved role and experience]. Show [evidence]. Do not add product features, results or personal experience. Return a script, shot list and any claim that still needs evidence.
Fill the brackets with information the campaign scope and product evidence support. Keep a record of the prompt, tool, source files and outputs. Review generated visuals and voice for product accuracy, identity use, captions and disclosures before asking the creator to approve the exact final version.
Separate concepts from execution changes
If two descriptions make the same argument, treat them as related executions even when their opening shots look different. If they answer different concerns, they may be separate concepts. These labels describe the team’s plan; they do not predict how a platform will classify ads.
Meta Blueprint teaches diversification by concept and motivator as well as by format. That supports a brief with more than a list of export sizes; it does not show that a particular batch will outperform current ads. Meta Blueprint on creative diversification. For interview material, see creator interview questions and script tracing.
Once the team chooses an argument, it can test a narrower execution change, such as starting with the bag or with the packed lunch. The creative-fatigue guide covers how many distinct ideas a budget can support. Label it as an opening test. A change in caption style does not by itself mean the client has approved a new argument or a new claim.
Plan a test the media team can interpret
Agree the learning question with the media buyer. For a comparison, write down what stays the same and what changes. For a wider creative exploration, record which features vary and limit the conclusion to the ads as a whole.
TikTok’s formal split test compares two versions, holds other variables steady and separates the audience into two groups. For more than two versions, plan pairwise rounds or confirm another test setup that supports the intended number of groups. If regular campaign delivery gives ads unequal spend, do not treat the dashboard ranking as a controlled comparison. Record delivery alongside results. TikTok split testing.
Keep a record through handoff
Give each file a concept ID, version ID, language and format. Record the tool, prompt, source files, creator scope, evidence checked and approval status. Note whether the file changes the argument, demonstration, performance or only the placement. Keep the exact-render sign-off in the creator approval workflow.
Keep rejected outputs out of the ready-to-run handoff, while retaining a production record when the team needs it for cost review. Send the media team only the approved files with current notes.
Review the set before generating more
Read scripts without thumbnails, then inspect visuals without sound. This helps the team see whether it changed the message, the appearance or both.
Ask whether the creator adds knowledge or performance that the ad needs. If the same evidence is missing from every version, another generation will not fix it. Return to the product facts, audience question or creator brief.
After the test, record which question the results can answer and which remain open. Use that record to choose the next task: develop a different argument, add evidence or refine an execution.
Connect this decision to the campaign
Set the agreed concept and output count in the campaign brief, including which changes need fresh review. If the aim is to reuse a willing partner’s context, assess the existing-creator scenario before promising no more filming.