Mike KarrCreative testing for ecommerce ads means comparing specific messages or executions to decide what to produce next. Start with one question, an approved control ad and a written decision rule. A folder of new videos is a production output; the test needs comparable delivery and a business outcome. This guide provides an original planning framework, not a record of campaigns we ran or a promise of sales lift.
Write the question before producing variations
Separate exploration from confirmation. Comparing a product demo with a presenter can help choose a concept to investigate, but changes several ingredients together. A controlled hook test asks a narrower question: does a different opening improve the chosen outcome when the rest stays fixed? Do not attribute a whole-concept result to its hook alone.
Copy this test brief: product and SKU / audience problem / hypothesis / control file / changed factor / fixed elements / primary outcome / diagnostic metrics / platform, placement and attribution window / review date and spend limit / decision rule / owner.
Example hypothesis: a visible organisation problem will bring more qualified shoppers than a question-led opening. Choose cost per purchase as the primary outcome if purchase tracking is reliable. Use click-through rate as a diagnostic. Define what improvement would matter for your economics and how much uncertainty you can accept before launch; no universal percentage makes a test valid.
Choose one factor you can interpret
| Factor | Change | Hold fixed | What the comparison can tell you |
|---|---|---|---|
| Hook | Opening line and first shot | Body, presenter, offer, duration and CTA | Whether the opening package deserves another purchase-focused test. |
| Angle | Organisation problem versus portability | Format, offer, audience and production quality | Which message package to investigate; not which individual sentence caused the result. |
| Presenter | Two suitable UGC/avatar presenters | Approved script, shot plan and destination | Whether this execution of the presenter changes response. Review delivery differences too. |
| Proof | Real compartment demo versus a static product view | Claim, hook, offer and CTA | Whether the evidence treatment helps. Use genuine approved product assets. |
| Format | Presenter versus product demonstration | Product facts, offer and destination | An exploratory concept comparison; several execution details change. |
Use this as a sequence of questions, not a requirement to launch every row. Begin with the uncertainty that blocks your next production decision. Save a reliable control file so later edits do not quietly change the baseline.
Worked example: two openings, one unchanged body
Hypothetical brief: a cable pouch with two compartments, sold to commuters. Approved assets show cables and a charger being placed inside. Those props are not included. We have not produced or advertised this example, and the feature description is invented for the exercise.
- A · problem opening · 0–3 seconds
- Show loose cables on a desk. Presenter: “Still searching for the right cable?”
- B · product opening · 0–3 seconds
- Show the open pouch. Presenter: “Two compartments for your everyday cables.”
- Shared body · 3–20 seconds
- Use identical approved footage of packing the pouch: “Put the cables in one compartment and the charger in the other. See the layout before you pack.”
- Shared close · 20–25 seconds
- Same product view, captions and CTA: “See the pouch details.” Link to the same product page.
This tests the opening package: both the first shot and line change. To isolate wording, keep the first shot identical instead. Keep the same voice, music, body, CTA, product page and export duration in both files. Avoid adding an unsupported convenience claim or a discount to only one version.
Produce and accept the actual test files
- Approve facts and the control
- Export controlled variants
- Compare compatible delivery
- Log a decision and retest
Kreopatra helps generate video ads from ready UGC/avatar templates, using a product URL or brief to develop scripts, previews and final rendered videos. Use it for the creative production step: select a suitable template, review the script and preview, then inspect the final download. Assemble separate real product footage in an editor when your proof requires it. Campaign experiments are set up in the ad platform.
- Verify the SKU, factual claims and rights to product footage. An AI presenter must not imply an invented customer experience.
- Watch both exports side by side. Check the unchanged sections, spoken delivery, captions, legibility and destination. If the tool regenerated a supposedly fixed scene, correct it or label the test more broadly.
- Name files before upload: pouch_hook_A_v1 and pouch_hook_B_v1. Store the approved brief, control and export versions together.
- Reject factual or production errors before paying for delivery. A polished preview is not approval of the final video.
Set up delivery and a review window
For a controlled comparison, use the platform’s experiment feature where available. TikTok describes split tests with two separate audience groups. Its variables guide allows one selected variable per split test. Its setup guide asks for a schedule of at least seven days. These are TikTok-specific instructions, checked October 28, 2026, not a universal minimum or a guarantee of a decisive result.
TikTok: split testing · TikTok: test variables · TikTok: setup and schedule
Before launch, record the objective, audience, placements, bidding setup, budget allocation, destination and attribution window. Keep those compatible across the comparison and check feature eligibility in the account. Ordinary campaign delivery may allocate impressions unevenly; that makes its observed results different from a controlled experiment. Choose a review date that accommodates the conversion delay, and record any changes or outages. Stop for broken tracking or incorrect claims; routine hourly winner-picking makes the interpretation less reliable.
Read purchases alongside attention
Illustrative arithmetic only: both versions spend $60 and receive 300 impressions in the same reporting window. A gets 30 outbound clicks and 3 purchases; B gets 45 clicks and 2 purchases. These are invented counts, not benchmark rates or results from Kreopatra.
| Metric | Version A | Version B |
|---|---|---|
| Outbound CTR = clicks ÷ impressions | 10% | 15% |
| Purchases | 3 | 2 |
| Cost per purchase = spend ÷ purchases | $20 | $30 |
| Purchase rate after click = purchases ÷ clicks | 10% | 4.4% |
B attracts more clicks, while A has the lower observed purchase cost. Neither has enough evidence here to declare a durable winner: the counts are tiny, uncertainty is unmeasured and conversion lag may change them. Check tracking and message-to-page fit, then use the planned review rule or run a follow-up. When purchases are zero, report zero purchases and undefined CPA rather than a fictional value. A proxy such as clicks can guide the next edit; it does not prove purchase lift.
Record a decision and the next test
| Observation | Decision to document | Next action |
|---|---|---|
| Comparable delivery and persuasive primary-outcome evidence | Candidate for confirmation | Repeat the promising concept before wider use; watch whether the effect holds. |
| Better attention, weaker purchase outcome | Message or destination hypothesis | Inspect the promise, product-page match and tracking; change one suspected cause. |
| Sparse or conflicting outcomes | Inconclusive | Reduce the number of variants, revisit the feasible budget and plan another test. |
| Tracking failure, unequal context or wrong export | Invalid for the intended question | Fix the problem and restart with a versioned control. |
Use a short result log: test ID / file versions / dates / actual spend and delivery / metric definitions and counts / attribution window / outages and edits / uncertainty / decision / next hypothesis. Save losing and inconclusive tests too. If performance later falls, inspect audience, frequency, placement mix, offer and website changes before blaming creative fatigue alone.
Questions about creative testing
How many creatives should I test at once?
Use the fewest variants that answer the current question with the traffic and review budget available. Two are enough for a focused A/B comparison. A large matrix is a production backlog, not an obligation to launch every combination.
How long should an ecommerce creative test run?
Plan around platform requirements, conversion delay and the evidence your decision needs. TikTok’s setup instruction above is platform-specific. A calendar deadline alone does not establish a winner.
Can a higher CTR identify the winning ad?
It identifies more clicks per impression under the measured conditions. For a purchase objective, review purchases, cost and uncertainty before choosing what to scale.
Can Kreopatra run or validate my ad test?
Use Kreopatra to prepare UGC/avatar video variants. Configure delivery and measurement in your ad platform, and keep a separate decision log. Generating an export does not validate its advertising performance.
Approve one product message, make two reviewable video variants, and take the finished files into a deliberately planned experiment.
Create a UGC/avatar video draft →